Paper deep dive

LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints

Ming Shin Cindy Sung, Duan Leijiahao, Haoxiang Wang, Boning Tan, Liangna Fu, Chao Tan, Yafu Li, Yulong Chen, Chaoqun Wan, Zhen Huang, Shuxian Liang, Xian-Sheng Hua

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 93

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Industrial manipulation in flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires robot policies to adapt to changing products and production conditions. Such settings impose industrial process constraints, including product variants, moving material flows, finite execution windows, and tight geometric tolerances. However, existing evaluation frameworks do not systematically capture these constraints when assessing generalist robot policies. To address this gap, we introduce LIBERO-Industry, a novel benchmark built on the widely used robot learning benchmark LIBERO, that translates the aforementioned industrial constraints into controllable task conditions and measurable success criteria. Specifically, LIBERO-Industry comprises three classic task scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. The suites contain 216 tasks in total that systematically vary object geometry, conveyor speed, interaction-window duration, and geometric acceptance region size. Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined. Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

90,120 characters extracted from source content.

Expand or collapse full text

LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints Ming Shin Cindy Sung Affiliation: Institute of AI for Engineering, Tongji University Duan Leijiahao Affiliation: Institute of AI for Engineering, Tongji University Haoxiang Wang Affiliation: Institute of AI for Engineering, Tongji University Boning Tan Affiliation: Institute of AI for Engineering, Tongji University Liangna Fu Affiliation: Institute of AI for Engineering, Tongji University Chao Tan Affiliation: Institute of AI for Engineering, Tongji University Yafu Li Affiliation: The Chinese University of Hong Kong Yulong Chen Affiliation: University of Aberdeen Affiliation: University of Cambridgehttps://IAIE-PhysicalAI.github.io/LIBERO-Industry Chaoqun Wan Affiliation: Institute of AI for Engineering, Tongji University Zhen Huang Affiliation: Institute of AI for Engineering, Tongji University Shuxian Liang Xian-Sheng Hua Affiliation: Institute of AI for Engineering, Tongji University Abstract Industrial manipulation in flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires robot policies to adapt to changing products and production conditions. Such settings impose industrial process constraints, including product variants, moving material flows, finite execution windows, and tight geometric tolerances. However, existing evaluation frameworks do not systematically capture these constraints when assessing generalist robot policies. To address this gap, we introduce LIBERO-Industry, a novel benchmark built on the widely used robot learning benchmark LIBERO, that translates the aforementioned industrial constraints into controllable task conditions and measurable success criteria. Specifically, LIBERO-Industry comprises three classic task scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. The suites contain 216 tasks in total that systematically vary object geometry, conveyor speed, interaction-window duration, and geometric acceptance region size. Extensive experiments across five generalist robot policies (including vision-language-action models and world-action models) show that performance progressively degrades as industrial constraints are introduced and combined. Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition. 11footnotetext: Corresponding author 1 Introduction Flexible manufacturing, e.g., high-mix, low-volume (HMLV) electronics assembly, requires frequent adaptation to changing products and production conditions, requiring substantial re-engineering effort on conventional automation (Gan et al., 2023; Schlenoff et al., 2021). Generalist robot policies, including Vision-Language-Action (VLA) models such as RT-2 (Brohan et al., 2023) and Octo (Ghosh et al., 2024), offer a learned alternative by conditioning robot actions on multimodal observations and language instructions across diverse tasks. However, industrial readiness requires more than task-level adaptability, particularly under the precision and operating conditions encountered in industrial settings (Li et al., 2025). Figure 1: LIBERO-Industry overview. Motivated by industrial process constraints, the benchmark evaluates generalist robot policies across three classic industrial scenarios. Industrial manipulation in flexible production is shaped by industrial constraints, introducing product variants, moving material flows, finite execution windows, and tight geometric tolerances, as illustrated in Figure 1. Product variants may require different setups and processing times (Gan et al., 2023), making frequent production changes harder to accommodate with fixed systems (Koren et al., 1999). Moving material flows impose finite time windows for manipulation (Islam et al., 2021), while parts must meet tight geometric tolerances despite manufacturing and perception errors (Jiang et al., 2022). Traditional automation works reliably when tasks and environments remain stable, but frequent changes often require reprogramming, recalibration, robot re-teaching, or changes to the workcell (Schlenoff et al., 2021; Gan et al., 2023) (detailed in Appendix A, Table 4). Together, these characteristics define the industrial process constraints that embodied manipulation evaluations should capture. However, existing embodied-manipulation and manufacturing-oriented benchmarks do not systematically evaluate generalist robot policy performance under industrial constraints, as summarized in Table 1. On one hand, general-purpose manipulation benchmarks, such as LIBERO, CALVIN, and Meta-World, provide standardized protocols for studying task learning, generalization, and language-conditioned control (Liu et al., 2023; Mees et al., 2022; Yu et al., 2019), yet largely abstract away the timing, motion, and placement requirements imposed by production processes. On the other hand, manufacturing-oriented research and benchmarks, including the Agile Robotics for Industrial Automation Competition (ARIAC), address robotic flexibility, dynamic production conditions, and operational performance (Harrison et al., 2018; Jiang et al., 2022) but do not comprehensively evaluate these industrial constraints or generalist robot policies. Consequently, how moving sources and destinations, finite execution windows, and tight geometric tolerances affect task success, throughput, and stage-wise failures remains underexplored, particularly as these requirements accumulate across multi-object workloads. To address this research gap, we introduce LIBERO-Industry, an extension of the widely used robot learning benchmark LIBERO (Liu et al., 2023) for evaluating generalist robot policies under industrial process constraints, that operationalizes these constraints as controllable task conditions and measurable success criteria. As shown in Figure 1, LIBERO-Industry comprises three task suites derived from typical industrial scenarios: Static-to-Dynamic (S2D) loading, Dynamic-to-Static (D2S) unloading, and Dynamic-to-Dynamic (D2D) transfer. These suites differ in whether workpiece sources and destinations are stationary or moving, capturing different forms of material-flow interaction. Each suite includes three progressive difficulty levels with controlled variations in workspace geometry, visual conditions, and task configuration, together with structured perturbations informed by industrial operating conditions and standards. Tasks couple spatial completion requirements with finite execution windows, allowing process constraints to be introduced systematically and in combination. We evaluate five generalist robot policies, including VLA models and a world-action model (WAM) in simulation, measuring task success, throughput, and stage-wise completion. We further assess real-world performance through physical D2S trials on an AUBO-i5H arm, comparing fine-tuning effectiveness with LIBERO-Industry and LIBERO under matched conditions. In summary, our main contributions are as follows: (a) Industrial Process Benchmark: We introduce LIBERO-Industry, a novel benchmark that systematically evaluates product variants, moving material flows, finite execution windows, and tight geometric tolerances. (b) Process-Constrained Task Design: We translate industrial process requirements as controllable task conditions, including conveyor velocity, bounded execution duration, and geometric completion requirement sizes, enabling their systematic evaluation individually and in combination. (c) Empirical Evaluation and Failure Diagnosis: We benchmark five generalist policies, including VLA models and a WAM, showing progressive performance degradation as industrial constraints are introduced and combined. Under a shared fine-tuning and evaluation protocol, the LIBERO-Industry co-training condition achieved higher observed target-transfer completion than the original-LIBERO condition. 2 Related Work Table 1: Comparison of recent embodied manipulation and manufacturing-oriented benchmarks across policy and industrial-process evaluation dimensions. ✓ denotes explicit evaluation, ∼ partial or indirect evaluation, and ✗ no evaluation. An execution time window denotes an operational constraint on task completion rather than a reported runtime metric. Dataset Domain Language-conditioned policy Product variants Material flow Execution time window Geom. tolerance Real-world validation General-purpose embodied-AI benchmarks LIBERO Household ✓ ∼ ✗ ✗ ✗ ✗ LIBERO-PLUS Household ✓ ∼ ✗ ✗ ✗ ✗ VLA-Arena Household ✓ ∼ ✗ ✗ ✗ ✗ LIBERO-X Household ✓ ∼ ✗ ✗ ✗ ✗ DynamicVLA Generalized ✓ ∼ ✗ ✗ ✗ ✓ DOMINO Generalized ✓ ∼ ✗ ✗ ✗ ✓ Manufacturing-oriented simulation benchmarks ARIAC 2025 EV battery ✗ ✓ ✓ ∼ ✗ ✗ AI-CPS Industry ✗ ∼ ✗ ✗ ✗ ✗ RoboBPP Logistics ✗ ✓ ✓ ✗ ∼ ∼ LIBERO-Industry (ours) Industry ✓ ✓ ✓ ✓ ✓ ✓ 2.1 Generalist Robot Policies Generalist robot policies map multimodal observations and language instructions to actions across diverse tasks. RT-2 and OpenVLA autoregressively decode discrete action tokens (Brohan et al., 2023; Kim et al., 2024), whereas π0 _0 and π0.5 _0.5 use flow-matching action heads to generate continuous action chunks (Black et al., 2024; Physical Intelligence, 2025b). Recent World Action Models (WAMs) couple future-state prediction with action generation. WorldVLA, DreamZero, and Cosmos Policy jointly predict future visual states and actions (Cen et al., 2025; Ye et al., 2026; Kim et al., 2026). VPP instead conditions action prediction on predictive visual representations; Fast-WAM uses video prediction for co-training but directly decodes actions at inference, while FLARE aligns policy representations with latent embeddings of future observations (Hu et al., 2024; Yuan et al., 2026; Zheng et al., 2025). However, their behavior under calibrated industrial operating conditions remains underexplored. 2.2 Generalist Robot Manipulation Benchmarks A variety of embodied manipulation benchmarks have been proposed to study multi-task learning, long-horizon control, and knowledge transfer. Meta-World targets multi-task and meta-reinforcement learning, CALVIN evaluates language-conditioned long-horizon execution, and LIBERO studies lifelong knowledge transfer (Yu et al., 2019; Mees et al., 2022; Liu et al., 2023). This line of work has expanded to robustness under object, scene, viewpoint, physical, and linguistic variations through LIBERO-PRO, LIBERO-PLUS, VLA-Arena, and LIBERO-X (Zhou et al., 2025; Fei et al., 2025; Zhang et al., 2025; Wang and others, 2026). More recent benchmarks, including DOMINO and the Dynamic Object Manipulation (DOM) benchmark introduced with DynamicVLA, consider dynamic disturbances and moving-object manipulation (Fang et al., 2026; Xie et al., 2026). Yet these benchmarks do not formulate production-line operating conditions as controllable evaluation variables for generalist policies. LIBERO-Industry addresses this gap through conveyor-driven material flows, finite execution windows, task-specific placement predicates, and repeated multi-object handling, together with process-informed perturbations derived from industrial operating conditions. 2.3 Industrial Process-Aware Evaluation A variety of manufacturing-oriented benchmarks have been proposed to evaluate production workflows, industrial manipulation, and robotic logistics. ARIAC assesses robotic agility under changing simulated production orders, AI-CPS evaluates industrial manipulation controllers in Isaac Sim, and RoboBPP studies physics-aware online bin packing using industrial data (National Institute of Standards and Technology, 2025; Zhou et al., 2024; Wang et al., 2025). Related manufacturing research further characterizes reconfigurability, material transport, cycle-time requirements, and pose accuracy (Gan et al., 2023; Groover, 2014; Mönch et al., 2021; International Organization for Standardization, 1998). Nevertheless, neither embodied manipulation benchmarks nor manufacturing-oriented evaluations systematically assess generalist policies under controlled production-line operating conditions. LIBERO-Industry targets this shared limitation by exposing domain-specific process conditions as controlled benchmark variables across standardized conveyor-transfer tasks. 3 LIBERO-Industry Benchmark Design The benchmark aims to evaluate how reliably and efficiently generalist robot policies perform material-transfer tasks in changing industrial workcells. To support comparable evaluation, each task instance is assigned explicit completion predicates and a predeclared episode horizon derived from its transport schedule. 3.1 Preliminaries and Task Interface We evaluate policies with different architectures and objectives through the language-conditioned observation–action interface used by LIBERO and CALVIN (Liu et al., 2023; Mees et al., 2022): ot∈,t∼πθ(⋅∣o≤t,ℓm),t∈,o_t , _t _θ(· o_≤ t, _m), _t , (1) where ℓm _m is the instruction for evaluation episode m, oto_t contains the benchmark-defined visual and proprioceptive observations, and tu_t is a robot action produced at control frequency fcf_c. The full simulator state sts_t is available to the evaluator for computing task predicates but need not be available to the policy. We denote the task instance for evaluation episode m by mT_m, with mC_m collecting its temporal opportunities and goal specification: m _m =⟨ℓm,μ0,m,m,ξmsrc,ξmdst,m⟩, = _m, _0,m,W_m, _m^src, _m^dst,C_m , (2) m _m =⟨(amj,Hmj)j=1Ntarget,m,m⟩. = \(a_mj,H_mj) \_j=1^N_target,m,G_m . Here, μ0,m _0,m is the initial-state specification; mW_m contains the target workpieces and any pre-existing destination workpieces; and ξmsrc _m^src and ξmdst _m^dst denote the source process and destination process, respectively, each of which may be static, continuously moving, or intermittent. For each of the Ntarget,mN_target,m target workpieces, indexed by j, amja_mj is its availability time and HmjH_mj is the end of its physical opportunity horizon. The set mG_m contains the goal propositions that define valid completion for the target workpieces and, when present, pre-existing destination workpieces. 3.2 Industrial Process Constraints Each task has two process-level conditions: transport determines how the source and destination evolve, while temporal opportunity determines when a transfer remains feasible. Transport condition. Under a continuous source process, target workpieces move at signed velocity VsV_s; under an intermittent destination process, the destination moves at signed velocity VdV_d between stationary service periods. These transport states change the workpiece or goal geometry over time. Speed is therefore an operating variable rather than a success metric. The continuous and intermittent modes follow established material-transport terminology (Groover, 2014). Temporal opportunity condition. For the continuous source process, target workpieces separated by center-to-center pitch dsrcd_src have transport-event interval τsrc=dsrc|Vs|. _src= d_src|V_s|. (3) For the pulse-indexed destination process, slot pitch sslots_slot and stationary service dwell TdwellT_dwell define the transport-event interval τdst=Tdwell+sslot|Vd|. _dst=T_dwell+ s_slot|V_d|. (4) These intervals define the physical transport schedule, while [amj,Hmj][a_mj,H_mj] bounds the opportunity during which workpiece j can be completed. For the selected source arrival intervals, see Appendix B.4. Figure 2: Task generation and scripted demonstration collection in LIBERO-Industry. BDDL specifies objects, process states, and goals; sampled layouts pass a feasibility gate or are resampled (top). A task-specific scripted expert attempts a rollout from a fixed initial state and seed (bottom). The diagram shows completion, kinematic-safety, and replay checks before accepted trajectories are stored as HDF5 demonstrations with provenance. Failed attempts can prompt replay-guided collector refinement. 3.3 Process-Aware Procedural Task Generation Figure 2 illustrates the task generation pipeline. While workspace support follows LIBERO’s region-based formulation (Liu et al., 2023), our generator additionally conditions object initialization and task timing on material-flow dynamics. Given an operation anchor, static fixtures are sampled with level-specific pose perturbations; moving sources are back-projected using transport direction, velocity, arrival interval, and offset; moving destinations set slot spacing from geometry and travel direction and service intervals from velocity and dwell (Appendix B.2). Each layout is serialized as BDDL with its initial predicates, transport conditions, instruction, and goals, and infeasible layouts are resampled before instantiation. 3.4 Task Suites and Difficulty Levels The source and destination transport states produce three non-static transfer suites, summarized in Table 2. Table 2: Task taxonomy and dataset composition of LIBERO-Industry across the three workflow suites. Full combinations are defined over object, conveyor speed, motion direction, and difficulty level. For D2S, four objects are selected as a stratified subset from six available objects. Suite Workflow logic Full combos Instantiated Train Test Obj. × Speed × Dir. × Lvl. D2S Conveyor Belt → Precision Slot 4 × 3 × 2 × 3 72 60 12 S2D Material Plate → Dynamic Slot 4 × 3 × 2 × 3 72 60 12 D2D Conveyor Belt → Dynamic Slot 4 × 3 × 2 × 3 72 60 12 Total 216 180 36 • Dynamic-to-Static (D2S) unloading: moving source → stationary fixture. The policy must acquire a continuously transported workpiece before its source opportunity ends and then satisfy the fixture-specific completion predicate. • Static-to-Dynamic (S2D) loading: stationary source → intermittently moving destination. The policy must place a workpiece during an available destination service period and maintain the required placement as the carrier resumes motion. • Dynamic-to-Dynamic (D2D) transfer: moving source → intermittently moving destination. The policy must satisfy both the source-acquisition opportunity and the destination-service opportunity before stable placement can be credited. The benchmark instantiates these suites with workpieces from multiple industrial domains and level-specific destination fixtures (for asset inventory see Appendix B.1). For task instance m, the benchmark generator converts the instantiated transport schedule and completion allowance into a rollout budget: Troll,m=tstart,m+Tevent,m+Tcomplete,m+tb,m,Kroll,m=⌈fcTroll,m⌉,T_roll,m=t_start,m+T_event,m+T_complete,m+t_b,m, K_roll,m= f_cT_roll,m , (5) where tstart,mt_start,m is the initialization or first-arrival offset, Tevent,mT_event,m is the time spanned by the relevant transport events, Tcomplete,mT_complete,m is the final manipulation or service allowance, and tb,mt_b,m is the terminal buffer. Appendix B.4 expands this expression for D2S, S2D, and D2D and reports the fixed allowances for every suite, level, and transport condition. These values are shared by all evaluated policies and are benchmark time limits rather than universal factory takt times. Each level preserves the same high-level transfer objective while jointly increasing completion precision and environmental variation: Level 1 pairs a broad tray region with the nominal workcell condition, Level 2 adds object-matched grooves with prescribed orientations and bounded perturbations, and Level 3 combines form-fitting inlays with broader perturbations. For moving destinations, success additionally requires carrier retention during transport, with perturbations specified using the pose-accuracy and pose-repeatability terminology of ISO 9283 (International Organization for Standardization, 1998) (see Secs. B.5.3, B.5.2, and B.6). 3.5 Evaluation Protocol and Metrics We report three metrics: binary task success rate and two partial completion rates. Primary metric: success rate (SR). For each g∈mg _m, let Em(g)∈0,1E_m(g)∈\0,1\ denote its Boolean evaluation at the predeclared rollout boundary Kroll,mK_roll,m. Episode success and Task Success Rate over M evaluation episodes are Ym=∏g∈mEm(g),SR=1M∑m=1MYm.Y_m= _g _mE_m(g), = 1M _m=1^MY_m. (6) Complementary metrics: partial completion rate (PCR). Partial Completion Rates measure progress toward the transfer objective. For episode m, let Ntransferred,mN_transferred,m denote the number of target workpieces credited by the transfer predicate, Nbase,mN_base,m the number of pre-existing destination workpieces, and Nslotted,mN_slotted,m the number of target and pre-existing workpieces satisfying their required terminal predicates. The episode-level scores are PCRT,m=Ntransferred,mNtarget,m,PCRS,m=Nslotted,mNtarget,m+Nbase,m.PCR_T,m= N_transferred,mN_target,m, _S,m= N_slotted,mN_target,m+N_base,m. (7) The first score isolates progress on target transfers, whereas the second also penalizes displacement of pre-existing destination workpieces. We report their episode means: PCRT=1M∑m=1MPCRT,m,PCRS=1M∑m=1MPCRS,m.PCR_T= 1M _m=1^MPCR_T,m, _S= 1M _m=1^MPCR_S,m. (8) 4 Experiments Following the experimental setup (Section 4.1), we evaluate overall benchmark performance (Section 4.2), analyze how performance changes as industrial stressors accumulate and where failures concentrate (Section 4.3), and assess whether co-training with LIBERO-Industry simulation data improves physical D2S performance and robustness over co-training with the original LIBERO data (Section 4.4). 4.1 Experimental Setup We follow the original LIBERO simulation settings except that each rollout is bounded by the task-specific horizon Kroll,mK_roll,m (Eq. (5)). We evaluate five publicly available LIBERO-finetuned checkpoints: π0.5 _0.5, OpenVLA-OFT, UniVLA, X-VLA, and Cosmos Policy. The deterministic evaluation set contains 36 tasks—12 each from S2D, D2S, and D2D—with four tasks per difficulty level in each suite and 10 initial states per task. We report SR, PCRTPCR_T, and PCRSPCR_S following Section 3.5 (details in Appendix C.1). Table 3: Evaluation-set results. Scores in %; 360 rollouts per policy. PCRSPCR_S includes initially valid destination workpieces that remain valid. Policy SR PCRTPCR_T PCRSPCR_S π0.5 _0.5 0.0 0.4 12.5 OpenVLA-OFT 0.0 0.0 11.2 UniVLA 0.0 0.0 12.2 X-VLA 0.0 0.0 11.9 Cosmos Policy 0.0 0.0 11.8 4.2 Benchmark Results Table 3 reports performance across the 36 held-out tasks. All five checkpoints were evaluated over 360 rollouts. None satisfied the complete-task success predicate, yielding an SR of 0.0% (0/360) for every policy. The suite-level breakdown likewise shows zero SR in S2D, D2S, and D2D (details in Appendix C.2). Among the evaluated policies, π0.5 _0.5 exhibits the strongest observed partial-completion profile. It is the only policy with nonzero target-transfer completion (PCRT=0.4%PCR_T=0.4\% overall, including 1.1% in S2D) and also records the highest terminal-slot occupancy (PCRS=12.5%PCR_S=12.5\%, compared with 11.2–12.2% for the remaining policies). The held-out BDDL specifications yield a mean initial-state slotting baseline of PCRSinit=12.31%PCR_S^init=12.31\% (12.3%12.3\% at the reporting precision of Table 3; see Appendix C.2). Thus, PCRSPCR_S values near this baseline largely reflect pre-existing destination workpieces that remain valid, whereas newly completed target transfers are isolated by PCRTPCR_T. Figure 3: Cumulative diagnostic stress progression for π0.5 _0.5. Each bar denotes complete-task success of the stress settings for n trials (D0–D1: n=200n=200; D2–D6: n=240n=240), and PCRTPCR_T denotes mean target-object transfer completion. Stars show mean policy steps over successful trajectories; D6 has no successful trajectories. The progression is diagnostic rather than causal. 4.3 Ablation and Findings Figure 3 traces the cumulative D0–D6 stress progression for π0.5 _0.5. SR declines from 97.0% at D0 to 43.8% at D4, then collapses to 2.1% at D5 and 0.0% at D6 as the task expands from one to multiple objects. Because each stage retains the preceding modifications, the progression is diagnostic rather than causal. Partial transfer persists at D5–D6 (PCRT=18.1%PCR_T=18.1\% and 11.5%11.5\%), while reach-stage failures account for 132/240 and 202/240 episodes, respectively, localizing the principal bottleneck to longer multi-object sequences (Appendix C.3; Appendix C.3.3). Figure 4: Real-world validation. A: AUBOi5H–CTAG2F90D D2S workcell. B: Per-step release-on-tray success for S1 and S2 at conveyor speeds of 0.03 and 0.05 m/s over 24 trials per condition; the table reports aggregate target-transfer completion as PCRTPCR_T out of 72 planned placements. C: Target-transfer completion (PCRTPCR_T) under five physical operating-condition variations, with exact values labeled. 4.4 Real-World Experiments We evaluate held-out physical D2S unloading on an AUBOi5H–CTAG2F90D workcell. Each trial requires the policy to transfer three workpieces sequentially from a moving conveyor into a stationary Level 1 tray. The evaluation uses four three-workpiece combinations and two held-out slot permutations per combination, yielding eight held-out object-order configurations. At each conveyor speed, each policy is evaluated over 24 trials, corresponding to 72 planned placements per policy–speed condition. The held-out combinations and permutations are disjoint from the physical training split (Appendix E.2; Table 18). We compare two sim–real co-training setups initialized from the same π0.5 _0.5 base checkpoint. S1 combines 1,724 original-LIBERO simulation demonstrations with 95 physical D2S demonstrations, whereas S2 combines 1,412 LIBERO-Industry simulation demonstrations with the same 95 physical demonstrations. Both use the same 9:1 frame-based simulation-to-real sampling ratio, 30,000 training steps, model configuration, and physical evaluation protocol. The comparison therefore isolates the simulation-data source and corpus size while holding the physical supervision and training procedure fixed (Appendix E.4; Table 20). Figure 4B reports target-transfer completion at conveyor speeds of 0.03 and 0.05 m/s. At 0.03 m/s, S2 achieves PCRT=34.7%PCR_T=34.7\% (25/72), compared with 26.4% (19/72) for S1; at 0.05 m/s, the corresponding values are 23.6% (17/72) and 4.2% (3/72). The fraction of trials completing the third placement is also higher for S2: 16.7% versus 4.2% at 0.03 m/s and 20.8% versus 0.0% at 0.05 m/s. Across both aggregate transfer completion and later placement steps, S2 maintains higher observed performance, with the largest separation at the faster conveyor speed. Figure 4C evaluates changes in illumination, object appearance, camera configuration, and table texture. S2 exceeds S1 across all five evaluated variations by 12.5–34.7 percentage points, with the largest difference under the camera variation (34.7% versus 0.0%). This consistent ordering indicates that the observed advantage of S2 extends beyond the nominal D2S configuration under the tested physical operating-condition changes. 5 Conclusion and future work We introduced LIBERO-Industry, a benchmark for evaluating generalist policies under industrial process constraints, including product variability, material flows, finite execution-time windows, and tight geometric tolerance. Built on LIBERO, the benchmark organizes these conditions into S2D , D2S and D2D task suites, with controlled variations in task configuration and operating conditions. Our results show performance degradation across the cumulative industrial stress progression and higher D2S manipulation ability and generalization for the LIBERO-Industry co-training condition than for the original-LIBERO condition. A key future direction is to determine whether generalist policies can retain flexibility without sacrificing industrial production rates; the benchmark’s finite opportunity windows abstract this requirement from heterogeneous factory needs. Future evaluation should move toward production-calibrated settings that relate cycle times to empirical operating rates and test repeated material arrivals, recovery events, and long-horizon throughput. Ultimately, industrial readiness should be assessed not only by task completion but by compatibility with diverse production processes. References Black et al. (2024) K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. π0 _0: a vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164. External Links: Document, Link Cited by: §2.1. Brohan et al. (2023) A. Brohan, N. Brown, J. Carbajal, Y. Chebotar, X. Chen, K. Choromanski, T. Ding, D. Welker, Z. Xu, F. Xia, et al. RT-2: vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818. Cited by: item 3, §1, §2.1. Cen et al. (2025) J. Cen, C. Yu, H. Yuan, Y. Jiang, S. Huang, J. Guo, X. Li, Y. Song, H. Luo, F. Wang, D. Zhao, and H. Chen WorldVLA: towards autoregressive action world model. arXiv preprint arXiv:2506.21539. External Links: Document Cited by: §2.1. Chi et al. (2023) C. Chi, S. Feng, Y. Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song Diffusion policy: visuomotor policy learning via action diffusion. In Robotics: Science and Systems (RSS), Cited by: item 2. Downs et al. (2022) L. Downs, A. Francis, N. Koenig, B. Kinman, R. Hickman, K. Reymann, T. B. McHugh, and V. Vanhoucke Google scanned objects: a high-quality dataset of 3d scanned household items. In IEEE International Conference on Robotics and Automation (ICRA), p. 2553–2560. External Links: Document Cited by: §B.1. Fang et al. (2026) H. Fang, S. Li, S. Wang, X. Xi, D. Liang, and X. Bai Towards generalizable robotic manipulation in dynamic environments. arXiv. External Links: Document Cited by: §2.2. Fei et al. (2025) S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, J. Fu, J. Gong, and X. Qiu LIBERO-plus: in-depth robustness analysis of vision-language-action models. arXiv. External Links: Document Cited by: §2.2. Gan et al. (2023) Z. L. Gan, S. N. Musa, and H. J. Yap A review of the high-mix, low-volume manufacturing industry. Applied Sciences 13 (3), p. 1687. External Links: Document Cited by: §1, §1, §2.3. Garrett et al. (2021) C. R. Garrett, R. Chitnis, R. Holladay, B. Kim, T. Silver, L. P. Kaelbling, and T. Lozano-Pérez Integrated task and motion planning. Annual Review of Control, Robotics, and Autonomous Systems 4 (1), p. 265–293. Cited by: item 1. Ghosh et al. (2024) D. Ghosh, H. R. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo, Y. L. Tan, L. Y. Chen, Q. Vuong, T. Xiao, P. R. Sanketi, D. Sadigh, C. Finn, and S. Levine Octo: an open-source generalist robot policy. In Proceedings of Robotics: Science and Systems, External Links: Document Cited by: §1. Groover (2014) M. P. Groover Automation, production systems, and computer-integrated manufacturing. 4th edition, Pearson, Upper Saddle River, NJ. Cited by: §2.3, §3.2. Haarnoja et al. (2018) T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International Conference on Machine Learning (ICML), p. 1861–1870. Cited by: item 2. Harrison et al. (2018) W. Harrison, A. Downs, and C. I. Schlenoff The agile robotics for industrial automation competition. AI Magazine 39 (4). External Links: Document Cited by: §1. Hogan (1985) N. Hogan Impedance control: an approach to manipulation: part i–theory. Journal of Dynamic Systems, Measurement, and Control 107 (1), p. 1–7. Cited by: item 1. Hu et al. (2024) Y. Hu, Y. Guo, P. Wang, X. Chen, Y. Wang, J. Zhang, K. Sreenath, C. Lu, and J. Chen Video prediction policy: a generalist robot policy with predictive visual representations. arXiv preprint arXiv:2412.14803. External Links: Document Cited by: §2.1. International Organization for Standardization (1998) International Organization for Standardization ISO 9283:1998 manipulating industrial robots — performance criteria and related test methods. Standard International Organization for Standardization, Geneva, Switzerland. External Links: Link Cited by: §B.6.1, §2.3, §3.4. International Organization for Standardization (2002) International Organization for Standardization ISO 8995-1:2002 lighting of work places — part 1: indoor. Standard International Organization for Standardization, Geneva, Switzerland. External Links: Link Cited by: §B.6.1. Islam et al. (2021) F. Islam, O. Salzman, A. Agarwal, and M. Likhachev Provably constant-time planning and replanning for real-time grasping objects off a conveyor belt. The International Journal of Robotics Research. External Links: Document Cited by: §1. Jiang et al. (2022) Y. Jiang, Z. Huang, B. Yang, and W. Yang A review of robotic assembly strategies for the full operation procedure: planning, execution and evaluation. Robotics and Computer-Integrated Manufacturing 78, p. 102366. External Links: Document Cited by: §1, §1. Kim et al. (2026) M. J. Kim, Y. Gao, T. Lin, Y. Lin, Y. Ge, G. Lam, P. Liang, S. Song, M. Liu, C. Finn, and J. Gu Cosmos Policy: fine-tuning video models for visuomotor control and planning. arXiv preprint arXiv:2601.16163. External Links: Document Cited by: §2.1. Kim et al. (2024) M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. OpenVLA: an open-source vision-language-action model. arXiv preprint arXiv:2406.09246. Cited by: item 3, §2.1. Koren et al. (1999) Y. Koren, U. Heisel, F. Jovane, T. Moriwaki, G. Pritschow, G. Ulsoy, and H. Van Brussel Reconfigurable manufacturing systems. CIRP Annals 48 (2), p. 527–540. External Links: Document Cited by: §1. Li et al. (2025) S. Li, Y. Chen, D. Li, S. Liu, D. Lan, Y. Liu, and Z. Pang Transferring vision-language-action models to industry applications: architectures, performance, and challenges. arXiv preprint arXiv:2509.23121. External Links: Document Cited by: §1. Linck and Cochran (1999) J. Linck and D. S. Cochran The importance of takt time in manufacturing system design. Technical report Technical Report 1999-01-1635, SAE International. External Links: Document Cited by: §B.4.1. Liu et al. (2023) B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone LIBERO: benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310. Cited by: §1, §1, §2.2, §3.1, §3.3. Lovano (2023) Lovano Conveyor belt assembly line 3d model. External Links: Link Cited by: §B.1. Mandlekar et al. (2023) A. Mandlekar, S. Nasiriany, B. Wen, I. Akinola, Y. Narang, L. Fan, Y. Zhu, and D. Fox MimicGen: a data generation system for scalable robot learning using human demonstrations. In Proceedings of the 7th Conference on Robot Learning, Proceedings of Machine Learning Research, Vol. 229, p. 1820–1864. External Links: Link Cited by: §D.1. Mees et al. (2022) O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard CALVIN: a benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks. IEEE Robotics and Automation Letters 7 (3), p. 7327–7334. External Links: Document Cited by: §1, §2.2, §3.1. Mönch et al. (2021) T. Mönch, A. Huchzermeier, and P. Bebersdorf Variable takt times in mixed-model assembly line balancing with random customisation. International Journal of Production Research 59 (15), p. 4670–4689. External Links: Document Cited by: §2.3. National Institute of Standards and Technology (2025) National Institute of Standards and Technology ARIAC 2025 documentation. External Links: Link Cited by: §2.3. Physical Intelligence (2025a) Physical Intelligence OpenPI: open-source models and packages for robotics. Note: https://github.com/Physical-Intelligence/openpiAccessed 2026-09-10 Cited by: §E.4. Physical Intelligence (2025b) Physical Intelligence π0.5 _0.5: a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054. External Links: Document, Link Cited by: §E.4, §2.1. ProductViz3D (2025) ProductViz3D Skincare cosmetics pack 3d model. External Links: Link Cited by: §B.1. Raibert and Craig (1981) M. H. Raibert and J. J. Craig Hybrid position/force control of manipulators. Journal of Dynamic Systems, Measurement, and Control 102 (2), p. 126–133. Cited by: item 1. Schlenoff et al. (2021) C. I. Schlenoff, Z. Kootbally, and E. Karpas Agile robotics for industrial applications: editorial. Robotics and Computer-Integrated Manufacturing 71, p. 102162. External Links: Document Cited by: §1, §1. Standardization Administration of China (2017) Standardization Administration of China GB/T 10595-2017 belt conveyors. Standard General Administration of Quality Supervision, Inspection and Quarantine of the People’s Republic of China, Beijing, China. External Links: Link Cited by: §B.6.1. Wang et al. (2026) G. Wang et al. LIBERO-x: robustness litmus for vision-language-action models. arXiv. External Links: Document Cited by: §2.2. Wang et al. (2025) Z. Wang, H. Zhao, J. Xu, S. Zhang, R. Hu, C. Zhu, Z. Zeng, W. Zhu, Z. Xiong, H. Yu, and K. Xu RoboBPP: benchmarking robotic online bin packing with physics-based simulation. External Links: Document, Link Cited by: §2.3. Whitney (1977) D. E. Whitney Force feedback control of manipulator fine motions. Journal of Dynamic Systems, Measurement, and Control 99 (2), p. 91–97. Cited by: item 1. Xie et al. (2026) H. Xie, B. Wen, J. Zheng, Z. Chen, F. Hong, H. Diao, and Z. Liu DynamicVLA: a vision-language-action model for dynamic object manipulation. arXiv. External Links: Document Cited by: §2.2. Ye et al. (2026) S. Ye, Y. Ge, K. Zheng, S. Gao, S. Yu, G. Kurian, S. Indupuru, Y. L. Tan, C. Zhu, J. Xiang, A. Malik, K. Lee, W. Liang, N. Ranawaka, J. Gu, Y. Xu, G. Wang, F. Hu, A. Narayan, J. Bjorck, J. Wang, G. Kim, D. Niu, R. Zheng, Y. Xie, J. Wu, Q. Wang, R. Julian, D. Xu, Y. Du, Y. Chebotar, S. Reed, J. Kautz, Y. Zhu, L. Fan, and J. Jang World action models are zero-shot policies. arXiv preprint arXiv:2602.15922. External Links: Document Cited by: item 3, §2.1. Yu et al. (2019) T. Yu, D. Quillen, Z. He, R. Julian, A. Narayan, H. Shively, A. Bellathur, K. Hausman, C. Finn, and S. Levine Meta-World: a benchmark and evaluation for multi-task and meta reinforcement learning. arXiv. External Links: Document Cited by: §1, §2.2. Yuan et al. (2026) T. Yuan, Z. Dong, Y. Liu, and H. Zhao Fast-WAM: do world action models need test-time future imagination?. arXiv preprint arXiv:2603.16666. External Links: Document Cited by: §2.1. Zakka (2022) K. Zakka obj2mjcf: a utility for converting composite obj files to a mujoco model. GitHub. Note: https://github.com/kevinzakka/obj2mjcf Cited by: §B.1. Zhang et al. (2025) B. Zhang, J. Li, J. Shen, Y. Zhang, Y. Cai, Y. Chen, J. Dai, J. Ji, and Y. Yang VLA-arena: an open-source framework for benchmarking vision-language-action models. arXiv. External Links: Document Cited by: §2.2. Zhao et al. (2023) T. Z. Zhao, V. Kumar, S. Levine, and C. Finn Learning fine-grained bimanual manipulation with low-cost hardware. In Robotics: Science and Systems (RSS), Cited by: item 2. Zheng et al. (2025) R. Zheng, J. Wang, S. Reed, J. Bjorck, Y. Fang, F. Hu, J. Jang, K. Kundalia, Z. Lin, L. Magne, A. Narayan, Y. L. Tan, G. Wang, Q. Wang, J. Xiang, Y. Xu, S. Ye, J. Kautz, F. Huang, Y. Zhu, and L. Fan FLARE: robot learning with implicit world modeling. arXiv preprint arXiv:2505.15659. External Links: Document Cited by: §2.1. Zhou et al. (2025) X. Zhou, Y. Xu, G. Tie, Y. Chen, G. Zhang, D. Chu, P. Zhou, and L. Sun LIBERO-pro: towards robust and fair evaluation of vision-language-action models beyond memorization. arXiv. External Links: Document Cited by: §2.2. Zhou et al. (2024) Z. Zhou, J. Song, X. Xie, Z. Shu, L. Ma, D. Liu, J. Yin, and S. See Towards building AI-CPS with NVIDIA Isaac Sim: an industrial benchmark and case study for robotics manipulation. In 2024 IEEE/ACM 46th International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP), p. 263–274. External Links: Document Cited by: §2.3. LIBERO-Industry: Benchmarking Robotic Manipulation under Industrial Process Constraints Appendix Contents Appendix A Control Paradigms and Technical Comparison This section provides a structured comparison of traditional, learning-based, and generalist robotic control paradigms. Table 4 outlines the trade-offs in model scale, data dependencies, and physical precision limits that motivate the transition toward high-fidelity simulation environments like LIBERO-Industry. Table 4: Comparative Analysis of Control Paradigms in Industrial Manipulation Tasks. Paradigm Representative Arch. / Algos Model Size & Freq. Data / Training Source Typical Precision & Repeatability‡ Traditional Rule-Driven FSM / BT; A* / TAMP; PID, Force, Hybrid Position–Force, Impedance No learned params; 1000–2000 Hz Zero data; physical models & expert kinematics Typically sub-millimeter repeatability in calibrated industrial setups Modular Perception-Action YOLOv8 / FastSAM + LLM-TAMP (MoveIt / OMPL) Perception: 100M–1B+ (<10<10 Hz); Planner: Det. (>200>200 Hz) Zero downstream trajectories; pre-trained vision & 3D CAD models Reported millimeter-scale alignment, depending on perception and calibration Imitation Learning (IL) ACT, Diffusion Policy, DP3 50M–100M params; 30–100 Hz inference Behavioral cloning from 50–100 expert demonstrations Sub-millimeter performance reported for structured task-specific settings Model-Free / Residual RL PPO, SAC, OEC-IRRL, RL-100 <20<20M params; >500>500 Hz Online/Offline RL via environment reward signals Task- and reward-dependent; sub-millimeter alignment reported in selected settings Generalist Robot Policies VLA models: RT-2, OpenVLA; generalist diffusion/transformer policies: Octo, RDT-1B 1B–7B+ params; 3–10 Hz Multi-task or multi-embodiment pre-training with language, image-goal, or action conditioning Centimeter-scale errors reported in selected precision-placement evaluations World Action Models (WAM) DreamZero, 1XWM, UniSim 14B+ params; 3–7 Hz closed-loop control Joint video-action diffusion / latent generative dynamics Task-dependent; no standardized physical-repeatability measure ‡ Precision and repeatability summarize representative reported results and depend on the robot, sensing and calibration pipeline, task geometry, and evaluation protocol. As detailed in Table 4, the landscape of robotic manipulation control can be broadly categorized into distinct evolutionary tiers, each exhibiting specific deployment bottlenecks when transitioned to High-Mix Low-Volume (HMLV) manufacturing: 1. Classical and Modular Control Stacks. Classical robot systems separate task sequencing, motion planning, trajectory generation, and feedback control. Finite-state machines and behavior trees encode process logic, while geometric or task-and-motion planners generate feasible motions. PID, force, hybrid position–force, and impedance controllers then regulate trajectory tracking and physical interaction [Garrett et al., 2021, Whitney, 1977, Raibert and Craig, 1981, Hogan, 1985]. These systems can provide high-rate feedback and repeatable execution under calibrated conditions, but their performance depends on accurate models, engineered task logic, and careful integration of perception, planning, and control. 2. Task-Specific Learning Policies. Imitation Learning methods such as ACT and Diffusion Policy learn observation-to-action mappings from expert demonstrations, whereas model-free Reinforcement Learning optimizes behavior through task-specific reward signals [Zhao et al., 2023, Chi et al., 2023, Haarnoja et al., 2018]. These approaches reduce the need to manually specify every motion or contact strategy and can represent temporally coordinated or multimodal behaviors. However, their performance remains closely tied to the training task, embodiment, observation–action interface, reward design, and coverage of demonstrations or environment interactions. 3. Generalist Robot Policies and World Action Models. Generalist robot policies use multi-task or multi-embodiment pretraining to support language-conditioned behavior and transfer across tasks. VLAs such as RT-2 and OpenVLA directly connect visual and language inputs to robot actions, while Octo and RDT-1B use generalist transformer or diffusion-policy formulations [Brohan et al., 2023, Kim et al., 2024]. World Action Models additionally predict future visual dynamics together with action sequences, providing an explicit generative representation of how the scene may evolve [Ye et al., 2026]. Their broader semantic and task coverage is promising, although runtime frequency, contact behavior, and physical precision continue to depend on the model architecture, inference hardware, action representation, and downstream adaptation. Appendix B Benchmark Design and Task Specification B.1 Task Suites, Assets, and Fixtures Each instantiated task is encoded as a BDDL specification containing its objects, initial conditions, language instruction, and goal propositions. The propositions in mG_m are grounded by the corresponding benchmark predicates. Figure 5 visualizes the suite–level matrix with one representative task from each combination. Figure 5: Representative LIBERO-Industry task instantiations. Rows show Dynamic-to-Static (D2S), Static-to-Dynamic (S2D), and Dynamic-to-Dynamic (D2D) transport. Columns progress from Level 1 sector trays to Level 2 object-specific grooves and Level 3 form-fitting inlays. Header and border colors identify difficulty levels and do not encode measured performance. We introduce a collection of industrial object assets spanning several manufacturing domains. The base models are adapted from publicly accessible and commercial asset repositories, including an industrial conveyor model [Lovano, 2023], Google Scanned Objects [Downs et al., 2022], and commercial design configurations [ProductViz3D, 2025]. Table 5 catalogs the validated workpiece inventory and the corresponding destination fixtures. Figure 6: Representative LIBERO-Industry workpiece and fixture assets. From left to right, the panels show six evaluation workpieces and their object-specific Level 2 grooves and Level 3 form-fitting inlays, spanning personal care, model figurines and vehicles, and consumer-electronics domains. We enforce a strict 1 cm geometric clearance rule across three difficulty tiers: • Sector Trays (Level 1): Domain-wide general receptacles dimensioned to fit the maximum bounding box of the largest workpiece within that industrial category plus a 1 cm margin. • Grooves (Level 2): Object-specific slots generated strictly from the bounding box dimensions of individual workpieces plus a 1 cm margin. • Form-Fitting Inlays (Level 3): Object-specific negative-space profiles modeled via proportional 1cm boolean mesh expansion in Blender. We process inlay configurations through the obj2mjcf utility using Volumetric Hierarchical Approximate Convex Decomposition (V-HACD) to decompose hollow geometries into convex sub-meshes Zakka [2022]. Table 5: Systematic Breakdown of Novel Evaluation Workpieces and Target Mating Fixtures Ordered by Insertion Difficulty Industrial Domain Base Workpiece Asset ID Complementary Fixtures Workspace conveyor_belt, conveyor_ghost Model Figurines mario, grimlock, bumblebee, toy_tray, & Vehicles dragon, mini_fire_engine, *_flat_groove, *_upright_groove, mini_roller, school_bus, *_v_groove, *_flat_inlay thomas, porter Personal Care cream_bottle, cream_jar, cosmetics_tray, & Cosmetics pump_bottle, serum_bottle *_flat_groove, *_upright_groove, *_flat_v_groove, *_v_groove, *_flat_inlay, *_upright_inlay Consumer mouse_naga, mouse_white_taipan, electronics_tray, Electronics mouse_black_taipan, headset, *_upright_groove, *_v_groove, keyboard *_upright_inlay * Denotes an object-specific fixture where * maps to the corresponding workpiece Asset ID. B.2 Catalog Composition and Spatial Distribution The BDDL catalog specifies target workpieces, non-target objects, and placement regions for each suite and level. Figure 7 compares the centers of its declared placement regions with LIBERO-Spatial; each point represents a region declaration, not a task or recorded demonstration. Figure 7: Declared placement-region centers. The dashed LIBERO-Spatial center range contains 510 of 528 Industry region declarations. Target and confounder composition. D2D. Level 1 contains four targets without confounders. Level 2 samples two or three targets from four source objects, leaving one or two source confounders. Level 3 contains two or three targets and three confounders distributed across source and destination locations. D2S. Level 1 contains three moving targets. Level 2 contains three targets and one source confounder. Level 3 contains two moving targets, one or two source confounders, and one prealigned destination object. S2D. Level 1 contains targets without confounders. Level 2 contains three targets with one or two source confounders. Level 3 contains three targets with upright/flat pose variation and up to three source and destination confounders in total. B.3 Language Instruction Construction and Perturbation Each task in LIBERO-Industry is associated with a fixed English instruction that serves as the language-conditioning input to the policy. The instructions were prepared offline through an LLM-assisted authoring workflow using structured task metadata, including the task suite, original manipulation objective, target objects, distractor objects, destination regions, and relevant intrinsic object properties. Controlled language perturbations were then introduced according to three level-specific rules while preserving the task-defining action, target, and destination. Table 6: Language-perturbation rules and representative instructions. Level Language-perturbation rule Representative instruction LV1 Lexical and syntactic paraphrasing of the original instruction using common synonyms. “Place the pump bottle onto the moving tray.” LV2 Replacement of explicit object names with intrinsic visual descriptions that distinguish the target from distractors. “Pick up the semi-transparent bottles with white pumps and place them into the moving conveyor slots.” LV3 Reformulation using industrial process and material-routing terminology while retaining the information required to identify the target and destination. “Route all white ambidextrous computer mice from the source conveyor into the moving destination slots.” Each finalized instruction is stored in libero_suite_language_map.py as a fixed mapping from BDDL task identifiers to instructions and is retrieved deterministically at runtime. B.4 Temporal Protocol and Operating Conditions B.4.1 Terminology and Scope Takt time is reserved for a demand-pacing requirement, τtakt=TavailableQdemand, _takt= T_availableQ_demand, (9) which relates available production time to required output [Linck and Cochran, 1999]. Because the benchmark does not prescribe a customer demand scenario, its temporal settings are reported as arrival intervals, service-opportunity intervals, evaluation deadlines, and episode horizons. These quantities are not interchangeable: an episode horizon limits recording or rollout duration, whereas an arrival interval or service-opportunity interval specifies the workload presented to the controller. Table 7: Temporal quantities used in LIBERO-Industry. Symbol Name Operational meaning τtakt _takt Demand takt Available production time per required product; used only when demand is specified. τsrc _src Source arrival interval Time between consecutive source workpieces becoming available. τdst _dst Destination interval Time between consecutive destination service opportunities on a pulse-indexed line. τd _d Evaluation deadline Latest admissible completion time for a scored workpiece. TrollT_roll Episode horizon Maximum simulated or recorded duration; not a production-rate claim. B.4.2 Suite-Specific Physical Timing For a continuous source conveyor with signed velocity VsV_s and consecutive workpiece pitch dsrcd_src, the source arrival interval is τsrc=dsrc|Vs|. _src= d_src|V_s|. (10) Equivalently, when an arrival interval is authored directly in the task definition, the corresponding pitch is dsrc=|Vs|τsrcd_src=|V_s| _src. For a pulse-indexed destination conveyor, let TdwellT_dwell be the stationary service phase and sslots_slot the distance advanced by one slot. Its service-opportunity interval is τdst=Tdwell+sslot|Vd|, _dst=T_dwell+ s_slot|V_d|, (11) where VdV_d is the signed destination transfer velocity. The implementation derives sslots_slot from the fixture extent along the travel axis plus the declared inter-fixture clearance. The evaluated release records freeze the fixture asset, travel axis, clearance, VdV_d, and TdwellT_dwell for every task. S2D timing is therefore determined by the destination mechanism, not by an assumed robot movement time. D2D tasks retain both source and destination constraints as the ordered pair i=(τsrc,i,τdst,i), τ_i=( _src,i, _dst,i), (12) rather than collapsing two asynchronous lines into a single scalar interval. Table 8 summarizes the parameters that must be exported with the benchmark release. Table 8: Timing provenance by task suite. Suite Source schedule Destination schedule Reported physical limits D2S VsV_s, dsrcd_src, τsrc _src, tofft_off Static τsrc _src S2D Static VdV_d, sslots_slot, TdwellT_dwell τdst _dst D2D VsV_s, dsrcd_src, τsrc _src, tofft_off VdV_d, sslots_slot, TdwellT_dwell τsrc _src, τdst _dst The task-specific benchmark execution-time allowances are reported in Table 9. Table 9: Suite-specific nominal benchmark takt and destination service-time settings. Suite–Level Takt @ V3/M3 Takt @ V5/M5 Takt @ V8/M8 Service Time D2D–LV1 5.0 s 6.0 s 6.0 s 1.5 s D2D–LV2 5.0 s 6.0 s 6.0 s 2.0 s D2D–LV3 5.0 s 6.0 s 6.0 s 2.5 s D2S–LV1 4.0 s 4.0 s 4.0 s – D2S–LV2 5.0 s 6.0 s 6.0 s – D2S–LV3 5.0 s 6.0 s 6.0 s – S2D–LV1 – – – 1.5 s S2D–LV2 – – – 2.0 s S2D–LV3 – – – 2.5 s Note. V3/M3, V5/M5, and V8/M8 denote paired positive and negative source-conveyor speeds of ±0.03± 0.03, ±0.05± 0.05, and ±0.08± 0.08 m/s, respectively. Each table entry reports the nominal benchmark takt assigned at that speed condition. B.4.3 Suite-Specific Episode Horizons Equation (5) expands according to the transport events encoded by each suite. Each expansion reports both the continuous rollout duration and the corresponding integer number of control steps. D2S. The horizon covers the source offset, all budgeted source-arrival intervals, one robot-cycle allowance, and the terminal buffer: TrollD2S T_roll^D2S =toff+(Ns−1)τsrc+tr+tb, =t_off+(N_s-1) _src+t_r+t_b, (13) KrollD2S K_roll^D2S =⌈fcTrollD2S⌉. = f_cT_roll^D2S . S2D. The destination schedule must accommodate the slower of the robot cycle and the destination service interval for every target: TrollS2D T_roll^S2D =ti+Ntmax⁡(tr,τdst)+tb, =t_i+N_t (t_r, _dst)+t_b, (14) KrollS2D K_roll^S2D =⌈fcTrollS2D⌉. = f_cT_roll^S2D . D2D. The horizon combines the source-arrival schedule with one post-arrival allowance, one destination service interval, and the terminal buffer: TrollD2D T_roll^D2D =toff+(Ns−1)τsrc+tx+τdst+tb, =t_off+(N_s-1) _src+t_x+ _dst+t_b, (15) KrollD2D K_roll^D2D =⌈fcTrollD2D⌉. = f_cT_roll^D2D . Here, NsN_s is the number of budgeted source positions; the generator uses the four-position source capacity for D2S and D2D Levels 2–3. NtN_t is the number of S2D targets. The BDDL value tofft_off is the time before the first source arrival, not a spatial offset. During initialization it is converted to an upstream displacement from the conveyor center as doff=|Vs|toffd_off=|V_s|t_off. The fixed values are toff=1.5st_off=1.5\,s, ti=1.0st_i=1.0\,s, tr=3.0st_r=3.0\,s, tx=1.5st_x=1.5\,s, tb=1.5st_b=1.5\,s, and fc=20Hzf_c=20\,Hz. The ceiling operation ensures that the discrete episode includes the complete continuous-time horizon. B.4.4 Evaluation Freeze and Throughput Reporting The comparison protocol freezes the task definition, timing condition, deadline grid, episode horizon, control frequency, random seeds, settling dwell, and success predicate before comparing policies. No timing value is retuned for an individual policy. Results report the complete success–deadline curve for each task, alongside the physical feasibility flags. The theoretical opportunity rate associated with a scalar interval τ is qopportunity=3600τopportunities/hour.q_opportunity= 3600τ /hour. (16) Consequently, intervals of 44, 55, and 66 s correspond to 900900, 720720, and 600600 theoretical opportunities/hour, respectively. They are not good-part throughput. Observed good-part throughput is instead qgood=3600NsuccessfulTobservedproducts/hour,q_good=3600\, N_successfulT_observed /hour, (17) and is reported as a benchmark-conditioned measurement rather than a universal HMLV production rate. B.5 Conveyor Mechanics and Success Predicates B.5.1 Simulation Mechanics and Transport Physics The simulated transport system maintains physical continuity by updating the state vectors of ungrasped workpieces at every simulation step. Let B∈(3)R_B (3) represent the rotation matrix of the conveyor mechanism in the world coordinate frame, ^local a_local the localized linear travel axis, and vbeltv_belt the scalar speed parameter. The instantaneous linear velocity update vector ˙Oxy(t) q_O_xy(t) applied to the horizontal plane of workpiece O is formulated as: ˙Oxy(t)=B⋅^local⋅vbeltif Φbelt(O)∧¬Ψgrasp(O)˙Oxy(t)otherwise q_O_xy(t)= casesR_B· a_local· v_belt&if _belt(O) _grasp(O)\\ q_O_xy(t)&otherwise cases (18) where the Boolean predicate Φbelt(O) _belt(O) indicates verified topological contact with the active conveyor mesh, and Ψgrasp(O) _grasp(O) evaluates whether the robot’s parallel jaw actuators have established an active, non-slip force closure on the workpiece geometry. Upon satisfying Φbelt(O) _belt(O), an instant kinematic anchoring constraint synchronizes the object’s base linear velocity directly to the transport vector, preventing non-linear drifting behavior outside explicit physics-driven slippage events. B.5.2 Kinematic Overrides for Conveyor and Assembly Physics To bypass simulation chattering and enforce high-rigidity industrial constraints, we apply explicit position and velocity overrides to dynamic objects at the MuJoCo runtime level. In discrete-time physics engines, relying solely on surface friction to transport complex-mesh workpieces at high velocities can cause unphysical micro-slippage and geometric interpenetration. To address these limitations, Algorithm 1 dictates the initial conveyor transport kinematics. Specifically, the initial phase utilizes an instant kinematic anchoring mechanism to synchronize the workpiece with the belt upon establishing continuous geometric contact. Algorithm 1 Conveyor Transport Kinematics (Starting Phase) 1: sim.data, target_vel, obj_name, belt_geom_ids 2: is_contact←_in_contact_with(sim,obj_geom_ids,belt_geom_ids)is\_contact←\_in\_contact\_with(sim,obj\_geom\_ids,belt\_geom\_ids) 3: if not is_contact then 4: return ⊳ Allow standard gravity and free rigid-body dynamics 5: end if 6: qpos_adr,qvel_adr←GetJointAddresses(sim,obj_body_id)qpos\_adr,qvel\_adr (sim,obj\_body\_id) 7: if obj_name∉_anchored_quatsobj\_name∉\_anchored\_quats then 8: _anchored_quats[obj_name]←[1.0,0.0,0.0,0.0]\_anchored\_quats[obj\_name]←[1.0,0.0,0.0,0.0] 9: _anchored_pos[obj_name]←sim.data.qpos[qpos_adr:qpos_adr+2]\_anchored\_pos[obj\_name] .data.qpos[qpos\_adr:qpos\_adr+2] 10: end if 11: ⊳ Enforce orientation lock to prevent mesh interpenetration jitter 12: sim.data.qpos[qpos_adr+3:qpos_adr+7]←_anchored_quats[obj_name]sim.data.qpos[qpos\_adr+3:qpos\_adr+7]←\_anchored\_quats[obj\_name] 13: ⊳ Override velocities based on belt target speed 14: for i∈0,1i∈\0,1\ do 15: if |target_vel[i]|<1e-5|target\_vel[i]|<1e-5 then 16: sim.data.qpos[qpos_adr+i]←_anchored_pos[obj_name][i]sim.data.qpos[qpos\_adr+i]←\_anchored\_pos[obj\_name][i] 17: end if 18: end for Following the initial transport phase, Algorithm 2 governs the terminal insertion constraints into the moving conveyor slots. Subsequently, this terminal phase employs a multi-stage velocity arrest mechanism (Levels 1 to 3) to lock the workpiece inside the slot geometry upon satisfying strict dot-product orientation thresholds. Algorithm 2 Multi-Stage Slot Arrest Kinematics (Ending Target) 1: sim.data, vel_vec, delta_pos, lock_level 2: obj_qpos_adr,obj_qvel_adr←GetJointAddresses(sim,obj_body_id)obj\_qpos\_adr,obj\_qvel\_adr (sim,obj\_body\_id) 3: if AlignmentThresholdMet()AlignmentThresholdMet() then 4: if lock_level==1lock\_level==1 then 5: ⊳ Level 1: Planar Lock Only (X, Y locked to conveyor; Z/Rotation free) 6: sim.data.qvel[obj_qvel_adr:obj_qvel_adr+2]←vel_vec[0:2]sim.data.qvel[obj\_qvel\_adr:obj\_qvel\_adr+2] \_vec[0:2] 7: sim.data.qpos[obj_qpos_adr:obj_qpos_adr+2]+=delta_pos[0:2]sim.data.qpos[obj\_qpos\_adr:obj\_qpos\_adr+2] +=delta\_pos[0:2] 8: else 9: ⊳ Level 2 & 3: Full 3D Spatial Lock 10: sim.data.qvel[obj_qvel_adr:obj_qvel_adr+3]←vel_vecsim.data.qvel[obj\_qvel\_adr:obj\_qvel\_adr+3] \_vec 11: sim.data.qpos[obj_qpos_adr:obj_qpos_adr+3]+=delta_possim.data.qpos[obj\_qpos\_adr:obj\_qpos\_adr+3] +=delta\_pos 12: sim.data.qvel[obj_qvel_adr+2]←0.0sim.data.qvel[obj\_qvel\_adr+2]← 0.0 ⊳ Arrest Z-axis gravity 13: end if 14: end if 15: ⊳ Update moving groove target position continually 16: sim.data.qpos[qpos_adr:qpos_adr+3]←exact_world_possim.data.qpos[qpos\_adr:qpos\_adr+3] \_world\_pos 17: sim.data.qpos[qpos_adr+3:qpos_adr+7]←world_groove_quat_wxyzsim.data.qpos[qpos\_adr+3:qpos\_adr+7] \_groove\_quat\_wxyz B.5.3 Predicate Scaling and Constraint Matrix To accommodate the progressive complexity of industrial workflows, the state evaluation predicates are stratified across three difficulty levels. As environmental variance increases, the task definitions impose stricter kinematic constraints and require higher degrees of freedom (DoF) for successful execution. Table 10: Kinematic Constraint Matrix by Predicate. A checkmark (✓) indicates the mechanism is strictly verified for task success, while a cross (×) indicates it is relaxed or not applicable. Predicate Gripper Released Topological Contact Vector Axis Alignment Velocity Sync ExactOn ✓ ✓ × × DynamicOn ✓ ✓ × ✓ AlignedOn ✓ ✓ ✓ × Routed ✓ ✓ ✓ ✓ The mapping of these target definition predicates across task paradigms is summarized in Table 11. Table 11: Target Definition Predicates and Kinematic Constraints by Difficulty Level. Difficulty Level Receptacle Type Static Target (D2S) Dynamic Target (S2D, D2D) Control Constraint L1 Tray ExactOn DynamicOn 3-DoF L2 Slot / V-Groove AlignedOn Routed 5-DoF L3 Inlay AlignedOn Routed 6-DoF B.6 Environmental Generalization Settings B.6.1 Standards and Reporting Basis The perturbation families are organized using selected industrial measurement and operating terminology. ISO 9283 provides the pose-accuracy and pose-repeatability framework used to describe bounded object, fixture, and camera-pose deviations [International Organization for Standardization, 1998]. ISO 8995-1 provides the workplace-lighting context for controlled illumination changes [International Organization for Standardization, 2002], while GB/T 10595 provides the operating context for the commanded belt-conveyor conditions [Standardization Administration of China, 2017]. These standards provide reporting context but neither prescribe nor validate the simulator-domain bounds below, which are fixed by the benchmark generator. B.6.2 Perturbation Specifications Level 1 is the nominal LIBERO-Industry condition. Levels 2 and 3 introduce moderate and broad generalization conditions, respectively. Workspace Geometry. For object initialization, Level 1 scales each declared planar placement interval by 0.50.5 and fixes yaw. Levels 2 and 3 scale the interval by 0.20.2 and sample yaw from ⁡(0,2π)U(0,2π). D2D fixture translations are bounded by ±2± 2 cm in Levels 1 and 2; Level 2 additionally applies a shared yaw perturbation of ±5∘± 5 . Level 3 applies cardinal transpositions, bounds translation to ±2± 2 cm in x and to ±2± 2 or ±5± 5 cm in y according to the transposition, and applies yaw perturbations of ±5∘± 5 or ±10∘± 10 . D2S Level 1 uses common fixture translations of up to ±5± 5 cm. Level 2 retains geometry-bounded translations of up to ±5± 5 cm and applies belt and groove yaw perturbations of ±10∘± 10 . Level 3 combines cardinal transpositions with common translations of up to ±2± 2 cm in x and ±4± 4 cm in y, 55 m local inlay jitter, and yaw perturbations of ±10∘± 10 . S2D uses common translations of up to ±5± 5 cm in all tiers, adds independent source and destination yaw perturbations of ±10∘± 10 at Level 2, and combines cardinal transpositions with ±20∘± 20 yaw perturbations at Level 3. Visual Observations. Table 12 reports the camera, illumination, and workbench parameters shared across D2D, D2S, and S2D tasks. Table 12: Visual-observation perturbations across the three generalization tiers. Angular bounds are in degrees; translation bounds are Euclidean norms. Configuration Level 1: nominal Level 2: moderate Level 3: broad Agent-view orbit Azimuth 00; elevation 38.92234038.922340; distance 1.1306511.130651 m Azimuth 45,315±5\45,315\± 5; elevation 38.922340±538.922340± 5; distance 1.130651±0.11.130651± 0.1 m Azimuth sampled from [0,135)∪(230,360)[0,135)∪(230,360); elevation 38.922340±1038.922340± 10; distance 1.130651±0.21.130651± 0.2 m Agent-view rotation Fixed Rotation norm ≤1≤ 1 Rotation norm ≤2≤ 2 Wrist camera Not used Translation norm ≤2.29≤ 2.29 m; rotation norm ≤1≤ 1 Translation norm ≤4.57≤ 4.57 m; rotation norm ≤2≤ 2 Illumination Neutral profile; scale 2.52.5; no cast shadows Neutral, warm, or cool profile; scale 2.5±0.52.5± 0.5; cast shadows Neutral, warm, or cool profile; scale 2.5±1.02.5± 1.0; cast shadows Workbench surface Green matte Green, blue, or black; matte or glossy Same six controlled styles as Level 2 The illumination entries are normalized MuJoCo renderer parameters rather than photometric measurements in lux. They therefore provide reproducible simulator-domain conditions without asserting photometric conformity. Transport Dynamics. The named conveyor-speed conditions are anchored at 0.030.03, 0.050.05, and 0.080.08 m/s. D2D samples magnitudes within three non-overlapping intervals: [0.0300,0.0399][0.0300,0.0399], [0.0400,0.0649][0.0400,0.0649], and [0.0650,0.0800][0.0650,0.0800] m/s. Level 1 uses equal signed source and destination speeds, Level 2 uses unequal same-direction speeds, and Level 3 uses opposite directions. D2D service times are 1.51.5, 2.02.0, and 2.52.5 s for Levels 1–3. For D2S, Level 1 applies ±5%± 5\% relative speed variation around the selected condition, Level 2 applies ±0.005± 0.005 m/s absolute variation, and Level 3 uses the selected condition without additional speed noise. S2D uses the same Level 1 and Level 2 rules; Level 3 samples speed magnitude uniformly from [0.03,0.08][0.03,0.08] m/s while retaining the task direction. S2D service times are 1.51.5, 2.02.0, and 2.52.5 s for Levels 1–3. B.6.3 Reproducibility and Configuration All scene perturbations are deterministic for the task seed. For each task, the generator serializes the camera, lighting, surface, reset, and rollout parameters. The record is stored in libero/libero/benchmark/industry_env_args.json. Suite-specific geometry, composition, and transport rules are defined in notebooks/industry/autogen/configs.py, perturbations.py, task_composition.py, and the D2D, D2S, and S2D generation scripts. Appendix C Evaluation Protocol and Extended Results C.1 Evaluated Models and Checkpoints We evaluate five publicly released policies fine-tuned on the original LIBERO benchmark, without further fine-tuning on LIBERO-Industry. This is a cross-benchmark transfer evaluation, not an evaluation of unfine-tuned base models. Table 13 identifies the checkpoint sources. The exact weight revisions and model-specific inference settings should be recorded in the released evaluation configurations. Table 13: Original-LIBERO-adapted checkpoints used for cross-benchmark evaluation. Policy Checkpoint identifier π0.5 _0.5 gs://openpi-assets/checkpoints/pi05_libero OpenVLA-OFT moojink/openvla-7b-oft-finetuned-libero-spatial-object-goal-10 UniVLA qwbu/univla-7b-224-sft-libero X-VLA 2toINF/X-VLA-Libero Cosmos Policy nvidia/Cosmos-Policy-LIBERO-Predict2-2B C.2 Evaluation-Set Breakdown Table 14 reports the initial-state slotting baseline, followed by the suite-level results in Table 15 for the same five checkpoints as Table 3. Table 14: BDDL-declared composition of the 36-task held-out evaluation split. NtargetN_target denotes goal workpieces requiring manipulation and excludes initially valid destination workpieces; NbaseN_base denotes pre-existing target confounders. Counts are computed from the four selected BDDLs in each suite–level. The final column is the macro-average of task-level ratios, not a ratio of pooled counts. Suite NtargetN_target (L1/L2/L3) NbaseN_base (L1/L2/L3) Count Nbase/(Ntarget+Nbase)N_base/(N_target+N_base) PCRSinitPCR_S^init (%) S2D 12/12/12=3612/12/12=36 0/0/6=60/0/6=6 6/426/42 10.83 D2S 12/12/8=3212/12/8=32 0/0/4=40/0/4=4 4/364/36 11.11 D2D 12/9/10=3112/9/10=31 0/0/8=80/0/8=8 8/398/39 15.00 Overall 36/33/30=9936/33/30=99 0/0/18=180/0/18=18 18/11718/117 12.31 Table 15: Evaluation-set results by suite. Scores in %; 120 rollouts per policy per suite. S2D D2S D2D Policy SR PCRTPCR_T PCRSPCR_S SR PCRTPCR_T PCRSPCR_S SR PCRTPCR_T PCRSPCR_S π0.5 _0.5 0.0 1.1 11.9 0.0 0.0 11.1 0.0 0.0 14.3 OpenVLA-OFT 0.0 0.0 9.1 0.0 0.0 11.0 0.0 0.0 13.4 UniVLA 0.0 0.0 10.8 0.0 0.0 11.1 0.0 0.0 14.7 X-VLA 0.0 0.0 10.8 0.0 0.0 11.1 0.0 0.0 13.8 Cosmos Policy 0.0 0.0 10.8 0.0 0.0 10.0 0.0 0.0 14.6 PCRSPCR_S includes initially valid destination workpieces (Section 3.5). C.3 Cumulative Diagnostic Stress Progression C.3.1 Stress-Progression Setup Unless otherwise noted, the diagnostic progression uses the same evaluation setup as Section 4.1. Table 16 records the progression-specific task coverage, rollout counts, and language conditions. Table 16: Task and language composition of the cumulative diagnostic stress progression. Stage Registered tasks Trials Language condition D0 ABLD000–009 10×20=20010× 20=200 Ten original spatial-relation prompts, such as “between the plate and the ramekin.” D1 ABLD100–109 10×20=20010× 20=200 Generic prompt: “pick up the akita black bowl and place it on the plate.” D2 ABLD200–211 12×20=24012× 20=240 Three singular D2D/D2S/S2D templates without a conveyor-source reference. D3 ABLD300–311 12×20=24012× 20=240 Singular family-specific templates aligned with the introduced conveyor geometry. D4 ABLD400–411 12×20=24012× 20=240 D2D and S2D introduce “moving tray”; D2S wording remains unchanged. D5 ABLD500–511 12×20=24012× 20=240 The D4 templates use plural object nouns. D6 ABLD600–611 12×20=24012× 20=240 The plural D5 instruction templates are retained. C.3.2 Trajectory-Level Failure Classification Figure 8 maps each rollout to the first unmet manipulation phase, from pre-reach policy freeze and reach failure through post-release goal failure. Figure 8: Trajectory-level failure classification. Each episode is evaluated along the reach–grasp–lift–transport–place sequence. The decision rules expose the operational thresholds used for stage-specific labels, while workspace violations, joint-limit saturation, and unmatched cases remain global or fallback outcomes. The resulting labels are used in a cumulative diagnostic stress progression, not a one-factor-at-a-time causal ablation. From D0 to D6, the sequence adds an industrial background, removes a visual anchor, replaces the target with industrial trays, introduces a conveyor and temporal dynamics, and increases the sequence from one to three objects. Adjacent stages may change more than one factor, so their differences localize breakdowns without isolating causes. The D0–D6 results aggregate 200 rollouts at D0–D1 and 240 rollouts at D2–D6, with 20 indexed episodes per registered task. Numeric evaluator seeds were not retained; accordingly, we report the observed proportions descriptively and do not claim a prespecified confidence-interval analysis. D6 has no successful trajectory, so its policy-step statistic is undefined. C.3.3 Representative Failure Trajectories To complement the aggregate failure proportions in Figure 3, Figure 9 shows representative frame sequences from four classified rollouts. Each row contains one trajectory from the frozen π0.5 _0.5 tracker run and reports the corresponding stage-level frequency. The examples illustrate the operational meaning of the tracker outcomes; they are not an exhaustive validation of all classified episodes. Failure phase Representative trajectory frames Tracker outcome t1t_1 t2t_2 t3t_3 Spatial / reach D5; ABLD503; episode 14 frame 0 frame 189 frame 399 Spatial approach failure 132/240 episodes (55.0%) at D5. Grasp / lift D3; ABLD306; episode 10 frame 42 frame 147 frame 399 Kinematic lift failure 9/240 episodes (3.8%) at D3. Dynamic transport D4; ABLD406; episode 11 frame 84 frame 168 frame 231 Dynamic slip Dropped midway; 35/240 episodes (14.6%) at D4. Placement D5; ABLD509; episode 15 frame 160 frame 200 frame 399 Post-release goal failure 28/240 episodes (11.7%) at D5. Figure 9: Representative failure trajectories. Each row shows three temporally ordered frames from one failed rollout, followed by the trajectory-level classifier output and its frequency within the corresponding diagnostic stage. The examples link the aggregate categories in Figure 3 to observable trajectory behavior and are illustrative rather than exhaustive. Appendix D Cross-Embodiment Conversion and Policy Adaptation Embodiment and task-space tracking. We instantiate an AUBO-i5H arm with a CTAG2F90D gripper, using the gripper’s grip_site as the tool center point (TCP) provided by the factory provider. Panda and AUBO environments share the task definition and named non-robot scene state, while the AUBO-i5H arm and controller are initialized natively. After settling, corresponding task-body poses must agree within 10−610^-6 m and 10−310^-3 degrees. Because the arms have seven and six joints, respectively, conversion tracks Cartesian references rather than copying joint coordinates. In the shared simulator world frame, the AUBO-i5H action is tA=[clip[−1,1]⁡([k⁡(t+1P−tA)/spkLog⁡(Rt+1P(RtA)⊤)/sR])gtP],u^A_t= bmatrixclip_[-1,1]\! ( bmatrixk(p^P_t+1-p^A_t)/s_p\\ kLog\! (R^P_t+1(R^A_t) )/s_R bmatrix )\\ g^P_t bmatrix, (19) where k=3k=3, sp=0.05s_p=0.05 m, sR=0.5s_R=0.5 rad, and LogLog returns a three-dimensional rotation vector. The target controller follows the next Panda TCP reference online while retaining the source gripper schedule. This aligns task-space references, not joint trajectories or pixelwise camera observations. (a) Panda setup (b) Panda wrist (c) AUBOi5H–CTAG2F90D setup (d) AUBO-i5H wrist Figure 10: Representative Lift-scene setup and wrist observations. The comparison checks robot–gripper composition and view coverage; it does not assume pixelwise camera equivalence. Policy interface and visual convention. The policy receives external and wrist RGB images and a seven-dimensional state comprising TCP position, axis–angle orientation, and one CTAG driver coordinate. Its seven-dimensional action comprises six relative operational-space controller commands and one gripper command; no additional delta-action transform is applied. Images are stored at 256×256256× 256 and resized to 224×224224× 224 by OpenPI. For the aligned simulation releases, export and inference both use the same 180180-degree rotation of raw simulator images. Since the recorded HDF5 images already contain a vertical flip, export applies a horizontal flip to obtain this convention. D.1 Trajectory Conversion and Recovery Motivation and scope. Motivated by MimicGen’s demonstrated transfer of the same Panda demonstrations to Sawyer, IIWA, and UR5e [Mandlekar et al., 2023], we bridge the Panda-to-AUBO embodiment gap with a MimicGen-based recovery procedure that transforms object-centric end-effector segments into AUBO reference frames and retains only trajectories that pass three out of three replays. Our local implementation follows MimicGen’s subtask decomposition and reference-frame transformation principles, but uses our own LIBERO and AUBO execution, validation, and data-materialization code rather than the released software. source demonstrations or AUBO phase seeds → subtask library → failed AUBO initial state → transform and execute → task gates and 3/3 replay Figure 11: Recovery pipeline. A candidate is added to the training dataset only after generation and replay validation. Recovery procedure. Gripper events and task predicates split successful source rollouts into object-centric segments. LIBERO-Object and LIBERO-Spatial use separate pick and place segments from successful Panda demonstrations, whereas LIBERO-Goal and LIBERO-10 use object, articulated-fixture, and push phases selected from successful AUBO phase seeds to preserve long-horizon task invariants. Given a source end-effector pose Te,tsW^W\!T^s_e,t and the source and query reference poses TrsW^W\!T^s_r and TrqW^W\!T^q_r, respectively, each segment is transformed as Te,tqW=TrqW(TrsW)−1Te,tsW,^W\!T^q_e,t=^W\!T^q_r (^W\!T^s_r )^-1^W\!T^s_e,t, (20) which preserves the end-effector motion relative to the live object or fixture. Each attempt restores the failed conversion’s exact AUBO initial state, bridges to the transformed segment by Cartesian interpolation, retains the demonstrated gripper schedule, and executes the poses with the AUBO operational-space controller. A candidate is retained only if its task predicates hold, its arrays are finite and actions bounded, its joint-margin and controller-clipping checks pass, terminal success persists, and its stored actions succeed in three replays from the same initial state and scene. Accepted recoveries retain their demonstration identity and provenance and are merged only with replay-verified raw conversions, as summarized in Table 17. Table 17: Final Panda-to-AUBO conversion and recovery yield. Raw retained demonstrations passed all three exact-action replays. Recovery includes both direct conversion failures and replay-rejected raw demonstrations. Suite Raw 3/3 retained Replay-rejected Direct raw failures Recovered 3/3 Final retained LIBERO-Object 425 4 71 75 500 LIBERO-Spatial 147 38 315 353 500 LIBERO-Goal 216 41 243 284 500 LIBERO-10 184 37 279 316 500 Total 972 120 908 1,028 2,000 Appendix E Physical System, Real-World Dataset, and Sim–Real Evaluation E.1 Physical Workcell and Sensor Setup The physical dynamic-to-static (D2S) workcell uses an AUBO-i5H arm with a CTAG2F90D gripper to transfer objects from a moving conveyor into stationary target fixtures. A Pico headset and controllers support teleoperated demonstration collection. An external HIKON camera and a wrist camera provide the two recorded RGB views (Figure 12); the policy’s state and action interface is specified in Appendix D. The target fixture changes from a tray in Level 1 to object-matched grooves in Level 2 and form-fitting inlays in Level 3. The robot and two-view policy interface are shared across levels while placement constraints change at the destination. HIKON camera AUBOi5H–CTAG2F90D workcell Pico equipment OOD objects Level 2 grooves Level 3 inlays Figure 12: Physical D2S workcell and fixtures. Left: AUBOi5H–CTAG2F90D workcell with conveyor, target tray, and marked HIKON camera. Right, reading order: Pico teleoperation equipment, example OOD objects, Level 2 grooves, and Level 3 inlays. E.2 Real-World Collection, Splits, and Quality Control The physical study tests D2S transfer in the same robot and camera workcell across three levels. Level 1 uses tray targets; Level 2 introduces an initial-region or target-region confounder and uses matching grooves; Level 3 combines initial- and target-region confounders with form-fitting inlays. The levels probe pickup under scene variation and increasingly constrained target placement within the same conveyor-to-stationary task structure. Because these factors change together, level differences do not isolate their individual effects. The physical collection design uses four three-object combinations. For each combination, four slot permutations are assigned to training and two disjoint permutations are reserved for held-out evaluation, as summarized in Table 18. A permutation specifies the left-to-right object order in the target slots. The planned training design contains 4combinations×4training permutations 4\ combinations× 4\ training permutations (21) ×2directions×3disturbance levels=96task configurations. × 2\ directions× 3\ disturbance levels=96\ task configurations. Conveyor speed is balanced across these configurations rather than fully crossed with every factor. A task configuration is distinct from a recorded trajectory; the accepted demonstration counts are given in Appendix E.3. Table 18: Predeclared real-world training and held-out slot permutations. Letters denote School Bus (A), Mini Roller (B), Mario (C), and Thomas (D). Combination Training permutations Held-out permutations ABC ACB, BAC, CAB, CBA ABC, BCA ABD ABD, BAD, BDA, DAB ADB, DBA ACD ACD, ADC, CDA, DCA CAD, DAC BCD BDC, CBD, DBC, DCB BCD, CDB The held-out permutations are reserved from the physical training split. E.3 Physical Demonstration Collection and Recorded Views Figure 13 shows the Pico-assisted collection procedure. Figure 14 shows paired camera observations stored in an accepted HDF5 demonstration. Figure 13: Pico-assisted physical demonstration collection. Six photographs illustrate the collection procedure: initial scene, approach, grasp, transfer, placement, and release for a Level 1 tray transfer. Figure 14: Recorded physical demonstration id_001. Stored language: “Retrieve the items from the conveyor and set them onto the plate.” Each column pairs the table-camera view (top) with the wrist-camera view (bottom) at the same observation index of the aligned 10-Hz HDF5 episode. Times are relative to its first observation. Accepted dataset. The accepted physical dataset contains 95 demonstrations and 41,257 frames. The accepted counts for Levels 1, 2, and 3 are 32, 32, and 31, respectively. Episode id_092 was excluded because its nearest wrist-camera frame failed the predeclared 25 ms synchronization threshold. Both co-training conditions use this same accepted release. E.4 Sim–Real Co-Training Configuration Training conditions. Both runs load the same local copy of the released π0.5 _0.5 base weights (pi05_base/params) [Physical Intelligence, 2025b] and use the OpenPI training implementation [Physical Intelligence, 2025a]. S1 combines AUBO-native original-LIBERO recovered-v2 simulation demonstrations with physical demonstrations. S2 replaces the simulation component with LIBERO-Industry demonstrations. Both conditions use the same 95 physical demonstrations. Table 19 summarizes the training sources. At physical evaluation, S1 and S2 face the same D2S workcell, object arrangements, target fixtures, and scoring rules. The comparison therefore tests the effect of changing the simulation training source under shared physical supervision; it does not attribute any difference to an individual confounder or fixture. Training source S1 S2 Simulation dataset Original-LIBERO recovered-v2 LIBERO-Industry Simulation episodes 1,724 1,412 Simulation frames 313,102 620,660 Physical episodes 95 95 Physical frames 41,257 41,257 Table 19: Training sources for the two co-training conditions. The physical dataset is shared, while the simulation source differs. Sampling and optimization. Sampling operates on LeRobot frames using a 9:1 simulation-to-real ratio. This ratio specifies the sampling distribution rather than the proportion of unique trajectories. Both runs are configured for 30,000 optimizer updates with a global batch size of 252. The shared model and optimization settings are listed in Table 20. This comparison matches the physical-data source, sampling weights, and configured optimization budget, but not the sizes of the unique simulation datasets. Setting Shared configuration Model π0.5 _0.5 Base checkpoint pi05_base State input Continuous Action horizon 10 Simulation-to-real sampling 90:10, frame-based Global batch size 252 Configured optimizer updates 30,000 Optimizer AdamW Gradient-norm clipping 1.0 Learning-rate schedule Cosine schedule with warmup Warmup steps 10,000 Peak learning rate 5×10−55× 10^-5 Decay learning rate 5×10−55× 10^-5 Scheduler decay steps 1,000,000 EMA decay 0.999 Training seed 42 Training hardware Six NVIDIA A100 GPUs FSDP devices 2 Data-loader workers 2 Checkpoint save interval 1,000 steps Table 20: Shared configuration specified by the S1 and S2 launch records. The scheduler’s decay-step parameter is distinct from the number of optimizer updates.

Instant research discovery