Paper deep dive
Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them
Carlota Parés-Morlans, Nils Kuhn, Isabel Liu, Alberta Longhini, Jeannette Bohg
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We address the problem of understanding when and why Vision-Language-Action models struggle with contact-rich manipulation tasks that require precise physical interaction. Prior work has primarily focused on addressing contact failures through force-augmented architectures and training-time regularizers, yet the root causes of these failures remain underexplored. We identify two distinct failure modes underlying this gap. Precision failures are rooted in a flow-matching policy training mismatch, and force failures arise from the distinctive structure of force signals. We address each failure mode with a targeted mechanism and combine them into FACT, which achieves 66% average success rate across five contact-rich tasks against 41% for the best prior baseline, in an evaluation spanning almost 2,500 real-world rollouts.
Tags
Links
- Source: https://arxiv.org/abs/2608.01402v1
- Canonical: https://arxiv.org/abs/2608.01402v1
Trouble viewing inline? Open PDF directly →
Full Text
52,047 characters extracted from source content.
Expand or collapse full text
Demystifying When and Why VLAs Fail in Contact-Rich Tasks and How to Fix Them Carlota Parés-Morlans, Nils Kuhn, Isabel Liu, Alberta Longhini, Jeannette Bohg Stanford University https://stanford-iprl-lab.github.io/fact/ Abstract We address the problem of understanding when and why Vision-Language-Action models struggle with contact-rich manipulation tasks that require precise physical interaction. Prior work has primarily focused on addressing contact failures through force-augmented architectures and training-time regularizers, yet the root causes of these failures remain underexplored. We identify two distinct failure modes underlying this gap. Precision failures are rooted in a flow-matching policy training mismatch, and force failures arise from the distinctive structure of force signals. We address each failure mode with a targeted mechanism and combine them into FACT, which achieves 66% average success rate across five contact-rich tasks against 41% for the best prior baseline, in an evaluation spanning almost 2,500 real-world rollouts. Keywords: contact-rich manipulation, force sensing, vision-language-action models, imitation learning 1 Introduction Vision-Language-Action (VLA) models have significantly advanced robotic manipulation, enabling policies that generalize across diverse tasks such as object retrieval and shirt folding [1, 2, 3, 4, 5, 6, 7]. Despite these advances, contact-rich tasks such as connector insertion and precision assembly remain an open challenge. Unlike free-space manipulation, these tasks require continuous force-regulated interaction where contact forces vary with part geometry, material properties, and surface compliance. At the same time, visual feedback often degrades due to self-occlusion at the moment when precise corrective actions are most critical. Standard training paradigms and architectural choices are ill-suited to these requirements, and current models continue to struggle in this regime. Figure 1: Contact-rich manipulation tasks from our experiments. These tasks involve sustained or repeated physical interaction, where success depends on simultaneously regulating both motion and contact forces measured by a wrist-mounted force/torque sensor. Plug insertion (top) demands sub-millimeter alignment of the connector under partial occlusion while managing contact until the connector is fully seated in the socket. Board erasing (bottom) requires maintaining consistent contact with the surface throughout the stroke without damaging the surface. A growing line of work attempts to close this contact-rich manipulation gap by augmenting VLA architectures with force feedback, either through dedicated fusion modules [8, 9, 10] or by using force signals as auxiliary supervision during training [11, 12]. These approaches share two implicit assumptions. First, they frame the contact-rich manipulation gap primarily as a force-sensing problem, leaving unexamined whether failures could stem from the training procedure itself rather than from missing sensory modalities. Second, when force is incorporated, it is usually appended alongside vision and proprioception, a design that does not exploit the properties of force signals. We diagnose why VLA policies often fail in contact-rich tasks and identify two distinct failure modes, precision and force failures. Precision failures occur because flow-matching training starves the low-noise denoising regime where contact demands small, accurate corrective actions. Force failures occur because force signals contain informative dynamics that are not captured when force is treated as just another input stream. To address both, we propose FACT, Force-Aware Contact-Rich Manipulation via Timestep Modulation, a method that targets the root cause of each failure mode. FACT introduces a targeted noise schedule that reallocates post-training signal toward the contact-correction regime, and a time-aware force injection mechanism that captures and exploits the distinctive properties of force signals. We evaluate FACT across five contact-rich tasks, two of which are illustrated in Figure 1, spanning both precision-driven and force-driven failure modes. FACT consistently outperforms force-augmented VLA baselines by more than 25 percentage points. Ablations show that each component contributes to these gains. The targeted noise schedule improves performance by up to 20 percentage points on precision-critical tasks, while time-aware force injection adds up to 22.5 percentage points on force-critical tasks. In summary, this paper makes the following contributions: • A principled decomposition of VLA failures in contact-rich tasks into two causally distinct modes, precision failures and force failures, together with root-cause analyses. • A targeted noise schedule that addresses precision failures by reallocating gradient signal to the contact-correction regime, yielding consistent gains across all flow-based VLA architectures tested, with no additional parameters, data, or architectural changes. • A time-aware force injection mechanism that addresses force failures by integrating force signals according to their distinctive properties. 2 Related Work 2.1 Contact-Rich Manipulation Contact-rich manipulation comprises tasks whose success depends not only on reaching a goal pose, but also on reasoning through physical interaction with the environment, including friction, compliance, jamming, and force regulation. This diversity motivates distinct sensing and control strategies across the literature. Classical approaches rely on haptic feedback alone, where passive mechanical compliance shapes the wrench response to misalignment [13], active force control regulates end-effector wrenches during contact [14, 15], and search strategies resolve position uncertainty through structured exploration [16]. Learning-based methods extend haptics-only control by learning contact-aware behaviors directly from interaction. For example, FORGE uses reinforcement learning to solve peg insertion, gear meshing, and nut threading from wrist force alone [17]. Adding vision unlocks spatial reasoning and generalization. Vision-force methods fuse wrist-mounted or fingertip force sensing with visual observations for insertion under occlusion, surface wiping, and force-sensitive manipulation [18, 19, 20, 21], and vision-tactile methods combine visual observations with distributed fingertip sensors for dexterous contact-rich tasks [22]. Combining vision with haptic sensing is non-trivial, as haptic signals are sparse and localized while vision is dense and structured, and naive fusion often overfits to the visual stream [23]. Our work builds on this fusion literature but focuses on exploiting the properties of force signals to propose a simple and effective mechanism for integrating them into vision-based models. 2.2 Force-augmented Vision-Language-Action Models Vision-language-action models [1, 2, 3, 4, 5, 6, 7] have rapidly become the dominant paradigm for generalist manipulation, yet they still underperform on contact-rich tasks. A growing line of work has sought to close this performance gap by augmenting VLAs with force feedback, addressing three distinct challenges. The first is ensuring force reliably shapes action selection during contact, with approaches ranging from force-aware routing after the VL encoder [8] to torque-history tokens [9] and per-layer cross-attention [10]. The second is action-space resolution, where fine-grained contact corrections require faster updates than a VLM backbone can provide, with approaches including decoupling a slow VLM backbone from a fast action module [10]. The third is modality imbalance, preventing force from being dominated by the visual stream, with approaches ranging from information bottlenecks [11] to training-time distillation [12]. In contrast to prior work that addresses contact failures through architectural force integration, we show that the training noise schedule and structural properties of force signals play an equally important role. We propose targeted mechanisms for each indentified failure mode and compare against representative force-augmented VLAs, ForceVLA and TA-VLA [8, 9]. 3 Preliminaries We introduce the flow-matching policy formulation and notation used throughout the paper, with emphasis on the role of the noise level τ and its sampling distribution. Flow matching [24] trains a network vθv_θ to predict a time-dependent velocity field that maps samples from a noise distribution to a target distribution. In visuomotor policy learning, the target is the conditional distribution of action chunks given the current observation o. Given a clean action chunk 0a^0 and noise ϵ∼(,I) ε (0,I), the rectified linear interpolant defines noisy actions τ=(1−τ)0+τϵa^τ=(1-τ)a^0+τ ε for τ∈[0,1]τ∈[0,1], with constant target velocity ϵ−0 ε-a^0. The network vθv_θ is trained to match this target, conditioned on the observation o and noise level τ, via the flow-matching objective ℒFM(θ)=0,ϵ,τ[‖vθ(,τ,τ)−(ϵ−0)‖22].L_FM(θ)=E_a^0, ε,τ [ \|v_θ(o,a^τ,τ)-( ε-a^0) \|_2^2 ]. (1) At inference time, integrating the learned field from τ=1τ=1 to τ=0τ=0 maps an initial Gaussian noise action chunk 1a^1 to a clean action chunk 0a^0. The training distribution of τ is set by a noise-level scheduler. Many standard choices, including the beta schedulers used in recent flow-based VLAs such as π0.5 _0.5 [2], concentrate probability mass on large τ values and therefore under-train the near-clean regime that governs fine corrective control. We focus our evaluation primarily on π0.5 _0.5, with additional results on π0 _0 [1] to test backbone generality (Section 7). 4 Diagnosing Failures of VLA Policies in Contact-Rich Tasks Compared to general manipulation tasks, contact-rich tasks require both fine-grained correction and reasoning about contact forces. Figure 2 illustrates these requirements through plug insertion, contrasting a successful execution with two distinct failure modes: precision failure and force failure. In what follows, we characterize each failure mode in detail and trace it to a specific property of flow-matching VLAs. (a) Success (b) Precision failure (c) Force failure Figure 2: VLA failure modes on plug insertion. Force magnitude |F||F| is shown over time. Dashed grey traces repeat the success profile for reference. (a) Success. The plug approaches the socket (|F|=0|F|=0), makes contact and explores alignment (|F|>0|F|>0), enters the socket (|F|≈0|F|≈0), and fully seats, producing the sharp force rise at t1t_1. (b) Precision failure. The plug is misaligned at the socket entry and force builds and saturates at t2t_2 as the plug presses against the rim. (c) Force failure. The plug enters the socket correctly but the policy stops short of full insertion and no seating force rise is detected at t3t_3. The two failures have distinct causes and require different corrective strategies. 4.1 Precision Failures At contact, two compounding effects hinder the sub-millimeter corrections required for success, leading to precision failures such as that shown in Figure 2(b) Delta collapse. Contact-rich manipulation exhibits a strong distributional shift in action magnitude. In free space, action deltas are large and variable, but during contact they collapse to near zero (see Figure A.3 in the Appendix for per-task delta profiles over time). This low-delta regime must be reproduced precisely at the moment visual feedback is least informative, as contact occludes the precise alignment error and the scene appears nearly static during sub-millimeter corrections. Training starvation. The flow-matching action head generates fine-grained corrections in the low-τ denoising regime, yet commonly used Beta noise schedules in VLAs such as π0 _0 [1], π0.5 _0.5 [2], and SmolVLA [6] allocate only 8.9% of gradient signal to τ<0.2τ<0.2, as shown in Figure 3(a). As a result, the contact-correction regime is starved of training signal. Jointly, delta collapse and training starvation reveal that the policy is undersupervised in the denoising regime where contact corrections are generated, an imbalance that is fully addressable through the noise schedule without changes to data or model architecture. 4.2 Force Failures Contact-rich tasks require policies to detect and react to forces, as illustrated in Figure 2(a). However, even when force sensing is available, force-augmented VLAs struggle to use it effectively, leading to failures like the one in Figure 2(c). This difficulty stems from three properties of force signals. (a) LN vs. Beta noise schedule (b) Force-augmented π0.5 _0.5 action expert Figure 3: FACT method overview. (a) Logit-Normal (LN) concentrates 6×6× more gradient signal in the contact-correction regime (τ<0.2τ<0.2, gray line) than the default Beta(1.5,1.0)(1.5,1.0) schedule. (b) Time-aware force injection uses the current reading ftf_t to modulate per-layer sensitivity via AdaRMSNorm, while the force history ft−H:t−1f_t-H:t-1 is prepended as tokens for temporal context. Contact sparsity. Force signals are approximately zero over most free-space timesteps and become informative only during short contact intervals (see Figure A.3 in the Appendix for per-task force profiles over time). Under naive concatenation, the objective is dominated by near-zero force samples, which biases gradients toward ignoring force input entirely. Temporal structure. Force provides information at multiple timescales. The instantaneous measurement encodes the current interaction state, whereas recent history encodes the dynamics that led to it, including transients and cumulative force buildup. Omitting either timescale discards task-relevant contact information. Sensitivity modulation. The influence of force on the policy should depend on contact state. In free space, force readings should be effectively ignored, whereas at contact even small deviations should trigger corrective actions. Collectively, these three properties motivate a force-injection mechanism that accounts for the imbalance of contact events, preserves the temporal structure of force signals, and uses force to modulate rather than simply augment the policy. 5 Addressing Failures of VLA Policies in Contact-Rich Tasks The failure modes identified in Section 4 motivate two targeted interventions. To address precision failures, we replace the commonly used Beta noise schedule with the Logit-Normal (LN) schedule during post-training, which reallocates training signal toward the low-noise regime that governs contact correction. To address force failures, we introduce a time-aware force injection mechanism that conditions the action generation module on contact state and a short force history. Both changes preserve the base VLA architecture, require no additional data, and add negligible inference-time overhead. 5.1 Fixing Precision Failures: Logit-Normal Noise Schedule We propose to improve the learning of sub-millimeter actions by reinforcing the low-noise, fine-correction regime. Specifically, during post-training we replace the Beta schedule with the Logit-Normal schedule, biasing the noise distribution towards the contact-rich regime. It’s distribution is given as: f(τ)=1s2π1τ(1−τ)exp(−(logit(τ)+m)22s2),logit(τ)=ln(τ1−τ).f_T(τ)= 1s 2π 1τ(1-τ) (- (logit(τ)+m )^22s^2 ), (τ)= ( τ1-τ ). (2) τ is sampled via the reparameterisation: τ=σ(s⋅z−m),z∼(0,1),τ=σ(s· z-m), z (0,1), (3) where σ is the sigmoid function. This adapts the LN schedule of Esser et al. [25], originally proposed for high-quality image generation, to the contact-rich manipulation setting with location parameter m=1.5m=1.5 instead of m=0m=0 to bias post-training toward the contact-correction regime. With this parameterization, LN allocates 6×6× more gradient signal to τ<0.2τ<0.2 than the Beta schedule, as illustrated in Figure 3(a). Critically, LN requires no changes to the model architecture, no additional training data, and no extra parameters, making it directly applicable to any flow-matching VLA. 5.2 Fixing Force Failures: Time-Aware Force Injection Our force injection approach, illustrated in Figure 3(b), is built around three design choices, each targeting one of the force failure properties identified in Section 4.2. Figure 4: VLA failures on two contact-rich tasks. Each row shows an approach frame (purple) followed by a success (green) and failure (red) frame with the corresponding force profiles. Vertical lines denote the approach event and the times at which success or failure is registered. In USB insertion, failure is caused by sub-millimeter pose misalignment. In key insertion, failure is caused by missing the hard-stop and continuing to push past the seating point. The key is shown unoccluded for clarity although the lock is fully occluded during execution. Contact state. Targeting sensitivity modulation, we inject the current force reading into the normalization layers of the action generation module to adapt the policy’s responsiveness to the contact state. The most recent sensor window t∈ℝHw×6f_t ^H_w× 6 is mean-pooled to a 6-dimensional summary f¯t f_t, passed through a two-layer MLP, and projected to a per-layer scale modulation: Δγ(f¯t)=Wγϕforce(f¯t)∈ℝd. γ( f_t)=W_γ\, _force( f_t) ^d. (4) This correction is added to the AdaRMS scale at every layer: hl=(γl(τ)+Δγ(f¯t))⋅RMSNorm(hl−1)+βl(τ)+hl−1gl(τ).h_l= ( _l(τ)+ γ( f_t) )·RMSNorm(h_l-1)+ _l(τ)+h_l-1\,g_l(τ). (5) Δγ γ is shared across all layers and modulates the sensitivity of every layer in a single forward pass. Contact history. Targeting temporal structure, the preceding H=30H=30 steps (2 s) of F/T readings t−H,…,t−1\f_t-H,…,f_t-1\ are each encoded independently by a shared causal encoder [26]. The resulting tokens are prepended to the input of the action generation module, enabling the policy to reason over the trajectory of contact. Contact gating. Targeting contact sparsity, a gradient gate with threshold δ blocks gradient through all force encoding components for steps where no contact is detected, preventing the force encoders from fitting to uninformative near-zero readings. Together, the LN noise schedule and the time-aware force injection form FACT, Force-Aware Contact-rich manipulation via Timestep Modulation. LN addresses the training imbalance that undersupervises the contact-correction regime, while time-aware force injection exploits the structural properties of contact signals. In Section 7, we validate the generality of LN when applied to existing force-augmented baselines [9, 8] and show that FACT transfers across flow-matching backbones. 6 Experimental Setup Table 1: FACT outperforms all baselines across five contact-rich tasks, with consistent gains on both precision- and force-critical tasks. Tasks from left to right: plug insertion, USB insertion, button push, board erasing, and key insertion. All Method ↑ [%] ↓p p ↑ [%] ↓p p ↑ [%] ↓p p ↑ [%] ↓p p ↑ [%] ↓p p ↑ [%] π0.5 _0.5 30.0 – 37.5 – 12.5 – 100.0 – 15.0 – 39.0 π0.5 _0.5 + LN 50.0 .055.055 47.5 .249.249 57.5 <.001<.001 87.5 1.001.00 37.5 .020.020 56.0 FACT 57.5 .012.012 47.5 .249.249 75.0 <.001<.001 90.0 1.001.00 60.0 <.001<.001 66.0 ForceVLAπ0.5_ _0.5 32.5 .500.500 37.5 .591.591 12.5 .631.631 77.5 1.001.00 42.5 .006.006 40.5 TA-VLAπ0.5_ _0.5 30.0 .596.596 25.0 .926.926 20.0 .273.273 97.5 1.001.00 15.0 .622.622 37.5 We describe the tasks, data collection, baselines, and evaluation protocol used in our experiments. Tasks We evaluate on five contact-rich tasks spanning both failure modes, with full descriptions and task illustrations provided in Appendix A. Plug insertion and USB insertion are precision-critical, demanding sub-millimeter corrections under partial occlusion. Key insertion, button push, and board erasing are force-critical. Key insertion is fully occluded and requires recognizing the hard-stop force signature as the key seats. Button push requires probing until a force threshold is reached, which locks the button in place. Board erasing requires maintaining consistent contact throughout the task. Representative episodes for USB and key insertion are included in Figure 4. Data Collection and Training All experiments are conducted on a Franka Research 3 arm equipped with a wrist-mounted Bota SensONE F/T sensor (see Appendix B for hardware details). For each task, we collect 100 teleoperated demonstrations using a Haply Inverse 3 haptic device that provides force feedback to the operator. At the start of each episode, the target position is sampled uniformly over a 32×2032×20 cm surface to encourage robustness, and the robot’s home position is sampled uniformly within a 55 cm cube. Visual and proprioceptive observations are recorded at 15 Hz and synchronized with F/T readings recorded at 400 Hz. All methods, including baselines, are fine-tuned from the pre-trained π0.5 _0.5 checkpoint using LoRA for 20,000 steps on the same demonstrations. Baselines We compare FACT against the standard π0.5 _0.5 policy and two force-conditioned baselines, ForceVLA [8] and TA-VLA [9]. Although ForceVLA and TA-VLA were originally introduced using the π0 _0 architecture, we re-implement both methods on top of π0.5 _0.5 to enable a controlled comparison under a shared policy backbone. We additionally report π0 _0 backbone results in Table 4 to enable direct comparison with the originally proposed baseline configurations. For a detailed comparison of FACT against baseline architectures please refer to Appendix E. Metrics and Evaluation Protocol Each method is evaluated on 40 independent rollouts per task, with target and home poses sampled from the same distributions used during training. Across all conditions, this yields nearly 2,500 evaluation rollouts in total. A rollout is considered successful if the task is completed within a 60 s timeout. We report success rate (SR %) as the fraction of successful rollouts and express differences in percentage points (p). Statistical significance is assessed using Fisher’s exact test relative to the π0.5 _0.5 baseline. 7 Results and Discussion Table 2: LN improves the performance of baselines with different force architectures built on the same π0.5 _0.5 backbone. Method ↑ [%] ↑ [%] ↑ [%] TA-VLA 30.0 15.0 20.0 + LN 47.5 32.5 42.5 ForceVLA 32.5 42.5 12.5 + LN 57.5 30.0 30.0 Table 3: Ablating force components from FACT shows all are necessary, with the largest drops on force-critical tasks. Method SR[%] SR[%] SR[%] FACT 57.5 60.0 75.0 w/o grad th 42.5 27.5 57.5 w/o curr. read 60.0 35.0 50.0 w/o history 30.0 20.0 12.5 Table 4: FACT applied to the π0 _0 backbone alongside prior baselines, showing that gains transfer across backbone versions. Method ↑ [%] ↑ [%] ↑ [%] π0 _0 32.5 45.0 47.5 FACTπ0_ _0 70.0 50.0 60.0 ForceVLAπ0_ _0 55.0 42.5 42.5 TA-VLAπ0_ _0 30.0 27.5 27.5 We structure results around five experimental questions, each targeting a specific hypothesis about the identified failure modes and the components designed to address them. Does LN improve flow-based VLAs performance on precision-critical tasks? The π0.5+ _0.5+LN row in Table 1 shows that replacing the Beta schedule with LN improves plug insertion by ++20 p (p=.055p=.055) and USB insertion by ∼+ +10 p, supporting our diagnosis that the default schedule starves the contact-correction regime on precision-critical tasks. LN gains are also observed on force-critical tasks. Button increases by ++45 p with p<.001p<.001 and key insertion ++22.5 p with p=.020p=.020. These gains require no additional data or parameters. Does time-aware force injection improve performance on force-critical tasks? Adding time-aware force injection on top of π0.5+ _0.5+LN (FACT row in Table 1) yields ++17.5 p on button push and ++22.5 p on key insertion, confirming that force reasoning at contact is essential for tasks where visual observation is ambiguous, with p<.001p<.001 for both against π0.5 _0.5. On precision-critical tasks, the added force signal provides no statistically significant benefit, indicating that plug and USB insertion fail due to insufficient denoising time rather than missing force feedback. Board erasing stands out as an outlier among force-critical tasks, with near-perfect success rates across all methods. Sustained surface contact is largely handled by the compliant operational-space controller (Appendix C), which maintains consistent contact force without requiring explicit force reasoning from the policy. The sparse failures observed are due to visual misalignment causing the eraser to cover only part of the mark rather than failing to maintain contact with the surface. How does each force component address the identified contact properties? Table 3 ablates each force component. Force history matters most, with button push dropping by −-62.5 p and key insertion by −-40 p, confirming that temporal integration is essential for force-critical tasks. The instantaneous reading plays a smaller role, except on key insertion (−-25 p) where the hard-stop peak provides a completion cue that history alone cannot supply. The gradient threshold proves necessary on force-critical tasks, where contact events carry the decisive completion signal. To verify these gains reflect genuine force exploitation rather than a training regularization, we replace F/T readings with Gaussian noise during training and evaluation. Full results are in Appendix D. Does LN recover baseline performance on precision-critical tasks? Table 2 shows that adding LN to either baseline yields consistent gains. This supports that the Beta schedule imbalance is a key bottleneck and that LN can be used as a drop-in fix for other flow-based VLAs without requiring any architectural changes. Do LN and time-aware force injection transfer to other VLA backbones like π0 _0? Table 4 evaluates FACTπ0_ _0 alongside the baselines on the π0 _0 backbone. FACTπ0_ _0 achieves 70.0 % on plug insertion, 50.0 % on key insertion, and 60.0 % on button push, outperforming both ForceVLA and TA-VLA. This shows that LN and time-aware force injection transfer across backbones. Comparing FACTπ0_ _0 against FACTπ0.5_ _0.5, π0.5 _0.5 is higher on force-critical tasks by over ++10 p, while π0 _0 is higher on the precision-critical task by ++12.5 p. The π0.5 _0.5 advantage on force-critical tasks may reflect its more explicit and pretrained timestep conditioning, which could make the force injection more effective at modulating the action expert’s sensitivity. The π0 _0 advantage on precision-critical tasks may come from differences in pretraining or state representation that improve fine corrections. Overall, FACTπ0.5_ _0.5 achieves 66.0 % success, outperforming ForceVLAπ0.5_ _0.5 (40.5 %) and TA-VLAπ0.5_ _0.5 (37.5 %) with consistent gains on both precision- and force-critical tasks, suggesting that addressing each failure mode independently leads to more consistent gains across contact-rich tasks. 8 Conclusion We present FACT, a method that addresses two causally independent failure modes of flow-based VLAs on contact-rich tasks. We identify precision failures and force failures as distinct root causes, and propose targeted solutions for each. The LN noise schedule reallocates post-training gradient signal to the low-noise denoising regime where sub-millimeter corrections occur, yielding consistent gains on precision-critical tasks without additional parameters or data. The time-aware force injection mechanism captures the sparsity and temporal structure of force signals, yielding large gains on force-critical tasks where visual feedback is insufficient. The two components are complementary. LN targets a training deficiency independent of sensing, while force injection targets a sensing deficiency. Ablations confirm that each component is causally independent and that all three force features are necessary. Both components transfer across backbone versions, suggesting they address general limitations of flow-based VLAs. Limitations. Our evaluation is conducted on a single robot platform with a fixed wrist-mounted F/T sensor, leaving open how well the force injection mechanism generalizes to robots with different kinematic structures or sensor placements. LN is specific to flow-matching action heads and does not directly apply to autoregressive or diffusion policies. Time-aware force injection is designed for action heads with transformer layers modulated via RMS scaling and would require adaptation for different architectures. Finally, our task set covers five scenarios and broader evaluation across more diverse geometries and material properties would strengthen the generality of our conclusions. Acknowledgments This work was supported in part by Agile Robotics. Carlota Parés-Morlans is supported by a graduate fellowship from Knight-Hennessy Scholars at Stanford University. Nils Kuhn is supported by scholarships from the Friedrich Ebert Foundation and the German Academic Exchange Service (DAAD). Alberta Longhini is supported by a Wallenberg–Bienenstock Postdoctoral Fellowship. We thank Hila Chefer for helpful discussions, and Michelle Yi for discussions on environment setup, data collection, and camera footage for figures. 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, 2024. Intelligence et al. [2025] P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al. π0.5 _0.5: a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025. Zitkovich et al. [2023] B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. In Conference on Robot Learning, pages 2165–2183. PMLR, 2023. Bjorck et al. [2025] J. Bjorck, F. Castañeda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y. Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025. 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, 2024. Shukor et al. [2025] M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. O’Neill et al. [2024] A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024. Yu et al. [2025] J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y. Song, P. Cai, et al. Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation. arXiv preprint arXiv:2505.22159, 2025. Zhang et al. [2025] Z. Zhang, H. Xu, Z. Yang, C. Yue, Z. Lin, H.-a. Gao, Z. Wang, and H. Zhao. TA-VLA: Elucidating the design space of torque-aware vision-language-action models. In 9th Conference on Robot Learning (CoRL), 2025. Li et al. [2026] Y. Li, P. Tang, W. Zhang, C. Zhu, Y. Duan, W. Shi, X. Zhang, Z. Yang, J. Ji, and Y. Zhang. FAVLA: A force-adaptive fast–slow VLA model for contact-rich robotic manipulation. arXiv preprint arXiv:2602.23648, 2026. Zhang et al. [2026] Y. Zhang, Y. Wang, X. Sun, K. Huang, Z. Xu, J. Ji, Z. Che, J. Tang, and J. Sun. Craft: Adapting vla models to contact-rich manipulation via force-aware curriculum fine-tuning. arXiv preprint arXiv:2602.12532, 2026. Zhao et al. [2026] R. Zhao, W. Wang, Y. Ma, X. Li, F. E. H. Tay, M. H. J. Ang, and H. Zhu. FD-VLA: Force-distilled vision-language-action model for contact-rich manipulation. arXiv preprint arXiv:2602.02142, 2026. Whitney [1982] D. E. Whitney. Quasi-static assembly of compliantly supported rigid parts. 1982. Khatib [1987] O. Khatib. A unified approach for motion and force control of robot manipulators: The operational space formulation. IEEE Journal on Robotics and Automation, 3(1):43–53, 1987. Lefebvre et al. [2005] T. Lefebvre, H. Bruyninckx, and J. De Schutter. Active compliant motion: A survey. Advanced Robotics, 19(5):479–520, 2005. Chhatpar and Branicky [2001] S. R. Chhatpar and M. S. Branicky. Search strategies for peg-in-hole assemblies with position uncertainty. In Proceedings of the 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2001. Noseworthy et al. [2025] M. Noseworthy, B. Tang, B. Wen, A. Handa, C. Kessens, N. Roy, D. Fox, F. Ramos, Y. Narang, and I. Akinola. Forge: Force-guided exploration for robust contact-rich manipulation under uncertainty. IEEE Robotics and Automation Letters, 2025. Lee et al. [2020] M. A. Lee, Y. Zhu, P. Zachares, M. Tan, K. Srinivasan, S. Savarese, L. Fei-Fei, A. Garg, and J. Bohg. Making sense of vision and touch: Learning multimodal representations for contact-rich tasks. IEEE Transactions on Robotics, 36(3):582–596, 2020. doi:10.1109/TRO.2019.2959445. He et al. [2025] Z. He, H. Fang, J. Chen, H.-S. Fang, and C. Lu. Foar: Force-aware reactive policy for contact-rich robotic manipulation. IEEE Robotics and Automation Letters, 2025. Li et al. [2025] Y. Li, Q. Lyu, J. Yang, Y. Salam, and W. Wang. A hybrid framework using diffusion policy and residual rl for force-sensitive robotic manipulation. IEEE Robotics and Automation Letters, 2025. Chen et al. [2025] C. Chen, Z. Yu, H. Choi, M. Cutkosky, and J. Bohg. Dexforce: Extracting force-informed actions from kinesthetic demonstrations for dexterous manipulation. IEEE Robotics and Automation Letters, 2025. Bi et al. [2025] J. Bi, K. Y. Ma, C. Hao, M. Z. Shou, and H. Soh. Vla-touch: Enhancing vision-language-action models with dual-level tactile feedback. arXiv preprint arXiv:2507.17294, 2025. Liu et al. [2025] J. J. Liu, Y. Li, K. Shaw, T. Tao, R. Salakhutdinov, and D. Pathak. Factr: Force-attending curriculum training for contact-rich policy learning. arXiv preprint arXiv:2502.17432, 2025. Lipman et al. [2022] Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. Esser et al. [2024] P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024. Bai et al. [2018] S. Bai, J. Z. Kolter, and V. Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018. Casiez et al. [2012] G. Casiez, N. Roussel, and D. Vogel. 1€ filter: a simple speed-based low-pass filter for noisy input in interactive systems. In Proceedings of the SIGCHI Conference on human factors in computing systems, pages 2527–2530, 2012. Appendix A Task Descriptions (a) (b) (c) (d) (e) (a) External (top) and wrist-mounted (bottom) camera views for each task at a representative contact moment. From left to right: (A) plug insertion, (A) USB insertion, (A) board erasing, (A) button push, (A) key insertion. We evaluate FACT on the five contact-rich manipulation tasks illustrated in Figure A. Plug insertion Illustrated in Figure A, a two-pin power plug must be inserted into an extension socket. Once the plug approaches the socket, the connector body occludes the holes, removing visual confirmation of alignment from the wrist camera. Seating the plug requires sustained force to overcome the friction fit, as shown in the force profile of Figure A.1. USB insertion Illustrated in Figure A, a USB-A connector must be inserted into a port. During insertion, the connector body occludes the port from the wrist-mounted camera, making visual alignment unreliable. Seating the connector requires sustained force to overcome the friction fit, as shown in the force profile of Figure A.1. Board eraser Illustrated in Figure A, a board eraser must be moved across a whiteboard surface to erase a marked region. The task requires maintaining consistent contact throughout the stroke. Too little force leaves residue, while too much risks skipping or damaging the surface. Force feedback is the primary signal for regulating contact quality, as shown in Figure A.1. Button push Illustrated in Figure A, a button must be pressed by probing until a force threshold is reached, which in turn locks it in place. Demonstrations were collected such that pressing with maximum force is not a viable strategy. The contact force shown in Figure A.1 provides the primary signal for determining when sufficient pressure has been applied, as vision alone is unreliable for detecting the locking event. Key insertion Illustrated in Figure A, a key must be inserted into a lock and seated at the correct depth, neither undershooting nor overshooting. To force reliance on force signals, demonstrations were collected using visually identical keys of different lengths, creating deliberate visual ambiguity. We additionally apply a Gaussian blur with σ=2σ=2 to the camera images for this task, further limiting the visual cues available for judging seating depth. The force profile shown in Figure A.1 provides the primary signal for detecting when the correct seating depth has been reached. Figure A.1: Environment state and force profile for five tasks. Columns show three representative episode moments followed by the full force profile. Purple marks an approaching step before contact, red marks the dominant failure point, and green marks success. (a) plug insertion, (b) USB insertion, (c) board erasing, (d) button push, and (e) key insertion. Key insertion images show the lock with an open back for visualization purposes. The left column shows the robot hardware setup: the yellow box highlights the Bota force/torque sensor mounted between the robot wrist and the Robotiq gripper, blue boxes mark the wrist-mounted camera and the external camera, and the purple box outlines the ferromagnetic board on which task objects are repositioned between trials. Appendix B Hardware Setup All experiments are conducted on a single Franka Research 3 arm equipped with a wrist-mounted Bota SensONE 6-axis force/torque (F/T) sensor. The full setup is illustrated in Figure A.1. The robot is controlled via Cartesian impedance control at 1 kHz. Futher details on the controller design are provided in Appendix C. Visual observations are provided by an external Realsense D435 camera and a wrist-mounted ZED Mini camera, both operating at 15 Hz. F/T data is acquired at 400 Hz and pre-processed with a one-euro filter [27]. Each control step therefore receives a window of Hw=27H_w=27 raw sensor readings (⌈400/15⌉=27 400/15 =27). All task objects were mounted on 3D-printed bases with embedded magnets, secured to a ferromagnetic board to prevent displacement during execution while allowing repositioning between trials. The plug insertion target was a commercial power strip extension held in a 3D-printed magnetic mount. The USB insertion used the connector from the NIST Robotic Assembly Board in a 3D-printed magnetic mount. The button push and key lock mechanisms were fully 3D-printed. For board erasing, the board was fixed and the drawn mark to be erased was varied between trials. A full bill of materials and 3D-printable STL files will be released alongside the code. Across all tasks, object positions were sampled uniformly within a 32×2032×20 cm workspace. Table A.1: Noise-substitution ablation replacing the measured F/T stream with i.i.d. Gaussian noise, preserving input dimensionality while removing task-dependent force information. FACT shows significant drops on all three tasks (p<.001p<.001), confirming genuine force exploitation. In contrast, ForceVLA shows a significant drop only on key insertion, and TA-VLA shows no meaningful dependence on F/T, suggesting their architectures do not reliably extract force information. Method ↑ [%] ↓p p ↑ [%] ↓p p ↑ [%] ↓p p TA-VLA 30.0 .306.306 15.0 .919.919 20.0 .273.273 TA-VLA force → noise 22.5 – 25.0 – 12.5 – ForceVLA 32.5 .824.824 42.5 .006.006 12.5 .826.826 ForceVLA force → noise 40.0 – 15.0 – 17.5 – FACT 57.5 .090.090 60.0 <.001<.001 75.0 <.001<.001 FACT force → noise 40.0 – 5.0 – 17.5 – Appendix C Controller Design We use a two-rate control architecture that decouples 15 Hz policy updates from a 1 kHz joint-torque control loop, enabling compliant behavior during contact via operational-space control [14, 18]. Let ∈ℝ3x ^3 denote end-effector position and R∈SO(3)R (3) its orientation. The policy outputs Cartesian end-effector displacement commands Δ∈ℝ3 ^3 and angle displacements Δα∈ℝ3 α ^3. The controller framework comprises three stages of trajectory generation, impedance control, and operational-space torque computation. Trajectory generation. At each policy step, the trajectory generator computes a desired end-effector pose desp_des from the current pose tp_t and the commanded displacements Δ and Δα α. It then interpolates a smooth trajectory ξt=k,k,kk=t+T _t=\p_k,v_k,a_k\_k=t^t+T of position, velocity, and acceleration at 1 kHz, bridging the gap between the low-bandwidth policy and the high-bandwidth controller. Impedance control. A Cartesian impedance PD controller tracks the interpolated trajectory and computes a task-space acceleration command: u=des−p(−des)−v(−des),a_u=a_des-k_p(x-x_des)-k_v(v-v_des), (6) where pk_p and vk_v are manually tuned stiffness and damping gains. Compliance during contact makes the robot safer and allows it to slide along surfaces under uncertainty, which is beneficial for contact-rich tasks. Operational-space torque computation. Using the known kinematic and dynamic model of the robot, we compute joint torques from Cartesian-space accelerations via the dynamically consistent operational-space formulation [14]. The task-space acceleration ua_u is first mapped to a task-space force via the operational-space inertia matrix Λ() (q), =Λ()u,f= (q)\,a_u, (7) and then to joint torques: u=J⊤(), τ_u=J (q)\,f, (8) where J()J(q) is the Jacobian at joint configuration q and Λ() (q) is the joint-space inertia matrix. Table A.2: Comparison of force-augmented VLA architectures. All methods use the π0.5 _0.5 backbone and are fine-tuned with LoRA. Parameter counts refer to added modules only. AE = action expert. FACT ForceVLA TA-VLA Force encoder Shared causal TCN + 2-layer MLP Linear proj. + LIMoE 2-layer MLP Injection point AE (AdaRMS + tokens) VLM/AE bridge (MoE) AE (1 token) Temporal encoding H=30H=30 windows of 27 steps (≈2≈2 s) None 10 frames (≈2≈2 s) Contact gating grad. threshold (δ=0.5δ=0.5 N) None None Added params ≈2.2≈2.2 M ≈45≈45 M ≈2.1≈2.1 M Appendix D Noise Ablation To test whether the F/T input provides task-relevant haptic information rather than merely acting as an auxiliary input channel that regularises learning, we perform the ablation study shown in Table A.1. Specifically, we replace the measured F/T stream with i.i.d. zero-mean Gaussian noise with unit standard deviation. This preserves the input dimensionality and network pathway associated with F/T while removing task-dependent force cues. To match inference-time conditions, policies are trained from scratch with Gaussian noise in place of the measured F/T readings. Replacing F/T with Gaussian noise yields no statistically significant change on precision-critical tasks including plug insertion and button push, suggesting that the corresponding gains are largely attributable to regularisation from the additional input rather than direct exploitation of force information. In contrast, on force-critical tasks, ForceVLA benefits from real F/T on key insertion, where success drops from 42.5 % to 15.0 % under noise replacement. TA-VLA shows no meaningful dependence on F/T on either force-critical task, which may indicate that its compressed single-token history is insufficient to capture the temporal structure required by these tasks. Appendix E Baseline Architecture Comparison Table A.2 summarises the architectural differences between FACT and the two re-implemented baselines. All three are built on the same π0.5 _0.5 backbone with identical LoRA adapters (≈50≈50 M trainable parameters), output delta end-effector pose actions, and are executed by the controller of Appendix C. “Added parameters” counts only the modules introduced on top of this shared base. Where the added parameters live. The three methods sit at very different points on the capacity-vs.-temporal-context spectrum. TA-VLA adds a 2-layer force MLP (6→d→d6→d→d, ≈2.1≈2.1 M) that encodes a history of force readings into a single token appended to the action-expert prefix. ForceVLA projects the current F/T reading through a single linear layer (6→d6→d, ≈14≈14 K) and routes it through a 4-expert LIMoE block placed between the VLM prefix and the action expert, with almost all of its ≈45≈45 M added parameters sitting in the LIMoE experts and gate rather than in the force encoder itself. FACT splits its ≈2.2≈2.2 M budget between a shared causal TCN (≈0.2≈0.2 M, 4 dilated blocks, hidden width 64) that ingests both the current 27-sample window and the H=30H=30 past windows, and an AdaRMS conditioning head (≈2.1≈2.1 M, two MLP layers and a zero-initialized γ-projection) that modulates the action-expert RMSNorm scales at every layer. Force signal. All three methods consume the same 6-D wrist F/T readings from the Bota SensONE, but differ in temporal context. ForceVLA uses only the single reading synchronized with the current vision and proprioception step. TA-VLA keeps a history of such readings, encoding ≈2≈2 s of past F/T into a single token. FACT augments the current 27-sample window with H=30H=30 past window summaries (≈2≈2 s of history at 15 Hz), which the shared causal TCN processes jointly. Action space. For a controlled comparison, all three methods are trained with delta end-effector pose actions executed by the controller of Appendix C, keeping the action interface and controller identical across all methods. π0 _0 timestep conditioning. The π0 _0 backbone results of Table 4 use a modified π0 _0 action expert. To enable more direct conditioning on the flow-matching timestep, we replace the original π0 _0 time-injection mechanism with the adaptive RMS normalization used in π0.5 _0.5. In the original π0 _0 architecture, the timestep embedding is concatenated with each noisy action embedding and passed through an MLP before entering the action expert. Consequently, timestep information is introduced only at the input of the transformer and must be propagated through the token representation across subsequent layers. In our modified architecture, the timestep is instead encoded into a global conditioning vector that modulates the RMS normalization layers throughout the action expert. Appendix F LN Parameter Exploration The LN schedule of Section 5 is controlled by its location parameter m, which determines how much post-training signal is allocated to the low-τ contact-correction regime. To justify our choice of m=1.5m=1.5, we sweep m∈−1.5,−0.5,0.5,1.5m∈\-1.5,-0.5,0.5,1.5\ at fixed s=1s=1, keeping all other training and evaluation settings identical. All runs use the full FACT model, varying only m. We run this sweep on plug insertion, a precision-critical task where sub-millimeter alignment makes the low-noise regime decisive and where the effect of the schedule is therefore most visible. Each configuration is evaluated over 40 rollout episodes. Figure 2(a) shows how m reshapes the sampling density. Recall that τ is drawn as τ=σ(s⋅z−m)τ=σ(s· z-m), so increasing m moves probability mass toward τ=0τ=0. Success rate follows this reallocation, as shown in Figure 2(b). Performance rises sharply once a substantial fraction of the signal reaches τ<0.2τ<0.2. (a) Logit-normal distribution of τ for different values of m. Larger m shifts training signal towards low noise τ<0.2τ<0.2. (b) Plug insertion success rate rises with post-training signal in the contact-correction regime. Figure A.2: LN parameter exploration on plug insertion. (a) The fraction of post-training signal falling in the contact-correction regime τ<0.2τ<0.2 for different values of m. (b) Success rate on plug insertion as a function of m. As larger m shifts the training signal towards low noise τ<0.2τ<0.2, success rate rises sharply. All runs use s=1s=1 and are otherwise identical, with 40 rollout episodes per configuration. Appendix G Per-Task End-Effector Deltas and Force Profiles End-effector position deltas and contact forces for each task are included in Figure A.3. These profiles illustrate the temporal structure of the contact phase for each task, which FACT’s architecture is designed to capture. In particular, the transition into contact is marked by shrinking position deltas as the end-effector approaches the target and rising contact forces as it makes contact. The vertical line in each plot marks the onset of the contact phase. (a) Plug insertion (b) USB insertion (c) Board erasing (d) Button push (e) Key insertion Figure A.3: Sample end-effector position deltas (top row) and contact forces (bottom row) for each task. The vertical line marks the onset of the contact phase.