Paper deep dive
LiLa-WAM: Lightweight Latent Reasoning World-Action Model for Robotic Manipulation
Fan Yang, Yuting Su, Xiaobo Wang, Yuncheng You, Fugui Fan, Yuting Wu, Minghui Wu, Chenxu Zhao, JiaHong Ning, Peiguang Jing
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:World-action modeling has emerged as a promising paradigm for robotic control, as it empowers models to go beyond reacting to observations and anticipate how a scene will evolve. However, existing WAMs often incur substantial computational overhead. Pixel-space methods often allocate substantial capacity to visual details that may not be directly relevant to control, while some latent-space methods require multi-stage training to construct the reasoning space. The resulting training cost can make such methods difficult to train under modest computational budgets. In this work, we propose LiLa-WAM, a lightweight world-action model that reasons about the future in a compact latent space and can be trained end-to-end on a single 24GB GPU. Its core design is a compact latent reasoning space jointly shaped by future-state prediction and action generation, which keeps the model lightweight while remaining well aligned with control. For task specification, we further propose the Visual Transition Token(VTT), a language-free task representation that encodes each task as a direction in visual feature space. Experiments on RoboTwin~2.0, LIBERO, and real-robot tasks demonstrate LiLa-WAM's effectiveness, achieving 90.48\% success across 50 RoboTwin tasks with single-GPU training.
Tags
Links
- Source: https://arxiv.org/abs/2608.03701v1
- Canonical: https://arxiv.org/abs/2608.03701v1
Trouble viewing inline? Open PDF directly →
Full Text
53,422 characters extracted from source content.
Expand or collapse full text
LILA-WAM: LIGHTWEIGHT LATENT REASONING WORLD-ACTION MODEL FOR ROBOTIC MANIPULATION Fan Yang 1 Yuting Su 1 Xiaobo Wang 2,5 Yuncheng You 1 Fugui Fan 1 Yuting Wu 4 Minghui Wu 3 Chenxu Zhao 3 Jiahong Ning 5 Peiguang Jing 1 1 Tianjin University 2 Shenzhen University of Advanced Technology 3 Mininglamp Technology 4 Ministry of Natural Resources Information Center 5 Sangfor Technologies Inc. August 5, 2026 ABSTRACT World-action modeling has emerged as a promising paradigm for robotic control, as it empowers models to go beyond reacting to observations and anticipate how a scene will evolve. However, existing WAMs often incur substantial computational overhead. Pixel-space methods often allocate substantial capacity to visual details that may not be directly relevant to control, while some latent- space methods require multi-stage training to construct the reasoning space. The resulting training cost can make such methods difficult to train under modest computational budgets. In this work, we propose LiLa-WAM, a lightweight world-action model that reasons about the future in a compact latent space and can be trained end-to-end on a single 24GB GPU. Its core design is a compact latent reasoning space jointly shaped by future-state prediction and action generation, which keeps the model lightweight while remaining well aligned with control. For task specification, we further propose the Visual Transition Token(VTT), a language-free task representation that encodes each task as a direction in visual feature space. Experiments on RoboTwin 2.0, LIBERO, and real-robot tasks demonstrate LiLa-WAM’s effectiveness, achieving 90.48% success across 50 RoboTwin tasks with single-GPU training. Code is available at https://github.com/teee000/LiLa-WAM. 1 Introduction World-Action Models (WAMs) [1–5] have emerged as a promising paradigm for robotic manipulation. By unifying predictive state modeling with action generation, WAMs enable the policy to anticipate how a scene will evolve under intervention, rather than merely react to current observations [6]. Such foresight can benefit robust control in manipulation tasks. Existing WAMs can be broadly grouped by the representation space in which the future is predicted. Pixel-space methods explicitly synthesize future observations as images or videos [7–11]. However, generating dense future frames, whether a single frame or a video sequence, tends to allocate substantial capacity to control-irrelevant details such as texture, lighting, and background, making both training and inference computationally expensive and memory- intensive. Latent-space methods instead predict future states in a latent representation space [12–16], which avoids such appearance redundancy. Nevertheless, a number of these methods construct their latent reasoning space through a multi-stage pipeline, where key components are often trained separately and subsequently integrated with the policy, further increasing the overall training burden. Despite this rapid progress, current WAMs remain expensive to train, limiting their accessibility under modest computational budgets. The primary source of the computational burden lies in the backbones. Current WAMs, regardless of their prediction space, are generally built upon large-scale pretrained models, whose capacity is largely devoted to functionalities beyond robotic control. For vision-language backbones, a considerable portion of parameters serves language modeling rather than physical interaction. For video generation backbones, substantial capacity is spent on synthesizing visual details that are not tightly coupled with action control. These observations suggest that compact, detail-rich visual backbones may offer a viable foundation for robot control, motivating a lightweight world-action model that relies solely on a visual backbone and receives task information from visual features rather than language. arXiv:2608.03701v1 [cs.RO] 4 Aug 2026 0.51248 Parameters (B, log scale) 65 70 75 80 85 90 95 Success Rate (%) ABot-M0 StarVLA 0 0.5 X-VLA Motus GigaWorld-Policy Being-H0.7 LiLa-WAM (Ours) (a) Comparison on RoboTwin 2.0 (50 tasks). 0.51248 Parameters (B, log scale) 75 80 85 90 95 100 Success Rate (%) OpenVLA OpenVLA-OFT StarVLA 0 GR00T-N1.6 SmolVLA EVO-1 LiLa-WAM (Ours) (b) Comparison on LIBERO. Figure 1: Average success rate versus model size on RoboTwin 2.0 and LIBERO, for methods with reported model sizes. LiLa-WAM achieves competitive average performance while using substantially fewer parameters than most compared methods, maintaining a lightweight design. In this work, we propose LiLa-WAM, a lightweight latent-reasoning world-action model for robotic manipulation. At its core, the Foresight-Aware Action Expert unifies future-state prediction and action generation within a single stream, allowing the two objectives to jointly shape a compact latent space. Beyond language instructions, recent works have explored specifying tasks through visual cues, such as interleaved image-text instructions or goal images [8]. Along this line, we take a further step toward simplicity by introducing the Visual Transition Token (VTT), a language-free task representation computed from the visual embeddings of task demonstrations. The VTT encodes each task as a transition direction in feature space to capture what the task changes about the scene. Compared with goal-image conditioning, it requires neither text nor a goal image at test time, providing a lightweight alternative for task specification. We evaluate LiLa-WAM on RoboTwin 2.0 (50 tasks with a single model), LIBERO, and real-robot tasks, and analyze its key design choices through extensive ablation studies and visualizations. Our contributions are summarized as follows: •We present LiLa-WAM, a lightweight world-action model that predicts future states and generates actions in a compact latent space, and can be trained end-to-end on a single 24 GB GPU. •We propose the Visual Transition Token (VTT), a language-free task representation that encodes each task as a transition direction in visual feature space, without requiring text or goal images at test time. •We empirically evaluate LiLa-WAM across simulation and real-world settings, where it remains competitive with larger policies. 2 Related Work 2.1 Vision-Language-Action Models Vision-language-action (VLA) models transfer web-scale vision-language pretraining into robot control [17–19]. Diffusion- or flow-matching-based policies [20–22] model continuous action chunks and have become the dominant architecture. A recent trend pursues lightweight VLAs that reduce backbone size and training cost [23–25]. Despite their differences, these models generally map the current observation directly to actions and remain largely reactive: how the scene will evolve under the robot’s actions is typically left unmodeled, and whether such foresight can emerge from demonstrations alone remains unclear. LiLa-WAM shares the lightweight pursuit but differs fundamentally in formulation. It couples action generation with explicit future-state prediction in a compact latent space, so that foresight becomes a training signal rather than an emergent by-product. 2.2 World-Action Models World-action models (WAMs) extend VLAs with the ability to predict future states alongside action generation. A first line operates in pixel space. Imagine-then-act methods [1,2,26–28] first generate future frames or videos and 2 Foresight-Aware Action Expert Feature Fusion Action Tokens DINOV3 Current Frame Future Frame LiLa-WAM DINOV3 Adapter Decoder Query Tokens Compact Latent Reasoning Tokens Cosine Loss Training Only VTT Foresight Tokens Figure 2: Overview of LiLa-WAM. The Foresight-Aware Action Expert unifies reasoning tokens, the VTT, propriocep- tive tokens, and noisy action tokens in a single stream, jointly producing the action velocity and foresight tokens, which are supervised in the feature space during training and discarded at inference. then recover actions through inverse dynamics or a future-conditioned policy, which decouples prediction from control at the cost of compounding errors and deployment latency. Joint video-action methods [5,10,29–32] instead model future frames and action chunks within a shared generative architecture, typically initialized from video-generation backbones such as WAN [33]. While effective, the appearance details these models learn to synthesize are often not strongly related to control, and training or adapting such generative backbones demands computational budgets far beyond a single GPU. A second line performs future prediction in latent space [13–15,34]. WoG [15] learns an action-oriented condition space by first grounding future observations into the action pipeline and then distilling them into the VLM. LaWAM [34] infers latent actions from visual transitions and trains a latent world model that decodes them into future features, which is integrated into a VLA policy in a second stage to provide latent visual subgoals. Being-H0.7 [16] shapes its latent space with a posterior branch that accesses future observations during training, relying on additional alignment losses to prevent feature collapse. These methods avoid pixel synthesis, but introduce their own complexities: WoG and LaWAM construct the reasoning space through two-stage pipelines that couple separately trained modules to the policy only afterwards, while Being-H0.7 requires auxiliary objectives beyond direct future supervision to keep the latent space well-behaved. LiLa-WAM also predicts the future in latent space, but integrates future-state prediction and action generation into a single stream trained end-to-end. 2.3 Visual Task Specification Beyond the dominant channel of language instructions, recent works have explored visual alternatives for task spec- ification [35–37]. Goal-image and subgoal-image conditioning [8,38] specify the task through a target visual state, yet require a concrete goal frame at test time. VIP [39] pretrains policies to interpret an instruction image depicting the desired interaction, and Interleave-VLA [40] shows that interleaved image-text instructions substantially improve out-of-distribution generalization over text-only conditioning, at the price of heavier multimodal inputs and instruction construction. Our Visual Transition Token takes a further step toward simplicity: it encodes the task as a transition direction in visual feature space computed offline from demonstrations, requiring neither text, nor paired instruction construction, nor a goal frame at deployment, while empirically concentrating on task-relevant objects. 3 Method 3.1 Problem Formulation We consider robot manipulation from visual observations. At time stept, the robot receives an image observation o t and a proprioceptive stateq t , and predicts an action chunka t:t+H = (a t ,a t+1 ,...,a t+H−1 )over horizonH. A standard vision-language-action policy models the conditional distributionp(a t:t+H | o t ,q t ,l), mapping the current observation and a language instructionldirectly to actions. Such a formulation focuses mainly on reacting to the current observation, with limited explicit modeling of how the scene will evolve once the robot acts. 3 World-action models address this limitation by modeling future states and actions jointly rather than actions alone. Following this principle, LiLa-WAM learns p θ a t:t+H , ˆ z t+∆ | o t ,q t ,c τ ,(1) where ˆ z t+∆ is the predicted compact latent of the future state after a look-ahead interval∆,θdenotes all trainable parameters, andc τ is a visual task cue that replaces the language instruction. Here the future is not represented by raw observations or pixel reconstructions, but by a compact latent produced within the action expert itself, which keeps Eq. 1 free from any heavy generative backbone. 3.2 Foresight-Aware Action Expert Motivation. The design of the action expert is guided by one goal: achieving lightweight yet foresight-aware world- action modeling. Existing WAMs fall short of this goal in different ways. Pixel-space methods synthesize future observations explicitly, spending substantial capacity on appearance details and relying on heavy generative backbones. Latent-space methods avoid pixel synthesis, but often construct their reasoning space through a multi-stage pipeline, where the predictive components are trained separately and coupled to the policy only afterwards. A deeper source of overhead lies in a mismatch between pretraining objectives and the demands of robot control. VLM-based VLA models inherit backbones pretrained with next-token prediction or image-text contrastive alignment on web-scale data [41, 42], objectives that favor high-level semantics over the fine-grained spatial and geometric cues that precise manipulation requires; video-generation-based WAMs likewise devote most of their parameters to modeling pixel-level dynamics rather than control-relevant state. In both cases, a substantial portion of the model’s capacity and computation may not be allocated to the cues most relevant to fine-grained control. We therefore build on DINOv3 [43], a self-supervised visual encoder whose dense pretraining yields patch features rich in fine-grained visual detail, and concentrate all trainable capacity in a single lightweight stream in which future-state prediction and action generation share and jointly shape one compact latent space. Architecture. We build LiLa-WAM on the frozen DINOv3 encoderVand compress its dense patch features into a small set of tokens with a query-based adapterA, Z v =A Q, ̃ V(o t ) ∈ R N q ×D ,(2) whereQ∈ R N q ×D is a set of learnable queries that cross-attends to the patch features,Z v is the compressed visual latent, ̃ V(o t )∈ R N p ×D denotes the fused multi-level features,N q is the number of queries, andDis the hidden width of the action expert. Prior studies on depth estimation and fine-grained recognition have shown that features from different backbone layers encode complementary information at different granularities [43–46]. Since precise action prediction likewise depends on such fine-grained cues, we exploit multi-level visual context before compression. Concretely, we concatenate the patch features of several layers ofValong the channel dimension and project the result back to the expert widthDwith a linear layer, so that the adapter receives a fusion of low-level spatial detail and high-level semantics at every patch. The adapterAstacks interleaved self-attention and cross-attention layers, and compresses the fused features intoN q tokens, yielding a compact latent space for foresight-aware reasoning. SinceN q is fixed regardless of the input resolution, the latent size is decoupled from the patch count, bounding the cost of all downstream attention. The action expertv θ is a stack ofLDiffusion-Transformer (DiT) blocks. At each flow-matching step, its input is a single token sequence Z a ; Z q ; c τ ; Z v ,(3) where Z a ∈ R H×D embeds the noised action chunk x s , Z q embeds the proprioceptive state q t , c τ is the VTT-derived task token, andZ v is the compressed visual latent from Eq. 2. After the final block, a linear head reads out the predicted velocity ˆ v s ∈ R H×d a at the action-token positions, while the observation-side outputs form the predicted future latent ˆ z t+∆ ∈ R N q ×D . Future-state prediction. During training, ˆ z t+∆ is decoded back to the feature space of the frozen encoder through a lightweight query-based (Q-Former-style) decoderD, and is supervised there against the patch features thatVextracts from the ground-truth future observationo t+∆ . In this way, future prediction serves as an auxiliary objective that shapes the shared latent, while no pixel decoder or heavy generative module is ever required. Because ˆ z t+∆ and the action velocity are read out from the same token sequence, gradients from the two objectives flow into one shared representation, which encourages the shared latent to remain compact and control-relevant. At inference, the expert inherently produces the future-aware latent together with actions in a single forward pass, and the decoder is discarded, introducing no extra test-time cost. 4 First Frames Last Frames VTT (a) Illustration of the VTT, computed as the mean difference between last- and first-frame embeddings. 00 adjust_bottle 01 grab_roller 02 hanging_mug 03 move_stapler_pad 04 open_microwave 05 press_stapler 06 scan_object 07 stack_blocks_two 08 stamp_seal 09 turn_switch (b) t-SNE of per-episode visual transition embeddings on 10 RoboTwin 2.0 tasks (clean and randomized). Figure 3: The Visual Transition Token and its embedding structure on RoboTwin 2.0. 3.3 Visual Transition Token Motivation. Manipulation policies typically specify the task with a language instruction, incurring text-processing overhead. We instead seek a task representation that is language-free, cheap to compute, and available without any goal frame during deployment. For a taskτ, letE τ be its set of demonstration episodes. For each episodee, we take the global image embeddings of its initial and final frames from the same frozen backbone, denotedg e 0 andg e T , and define the Visual Transition Token as their mean difference over the task, u τ = 1 |E τ | X e∈E τ g e T − g e 0 ,(4) where|E τ |is the number of episodes of taskτ. The VTT encodes the task as a single transition direction in feature space, i.e., what the task changes about the scene. The VTT is projected to the expert width with a lightweight MLP and injected into Eq. 3 as the task tokenc τ , a fixed per-task vector that requires neither text nor a goal frame at test time. 3.4 Training Objective LiLa-WAM is trained end-to-end with two complementary objectives: a flow-matching lossL fm for action-chunk prediction and a foresight lossL f for future-state prediction, L =L fm + λ f L f ,(5) where λ f balances the two terms. For action prediction, we adopt conditional flow matching. Letx 0 ∼N (0,I)be a noise sample,x 1 the ground-truth action chunk, ands∈ [0, 1]the flow time. We form the linear interpolationx s = (1− s)x 0 + sx 1 with the constant target velocity v ⋆ = x 1 − x 0 , and train the expert v θ to predict this velocity, L fm = E x 0 ,s ˆ v s − v ⋆ 2 2 ,(6) where the conditioning setC =o t ,q t ,c τ enters through the token sequence in Eq. 3, and the velocity is read out at the output positions corresponding to the action tokens Z a . Future-state prediction is supervised in the DINOv3 feature space. Specifically, a lightweight Q-Former-style decoder Dmaps the predicted future latent ˆ z t+∆ to the patch-feature space of the frozen encoder. The foresight lossL f is computed as a per-token cosine loss between the decoded features and the patch features extracted byVfrom the ground-truth future observation o t+∆ . The decoder is used only during training and is discarded at inference. 5 MethodParams Clean Random ABot-M04B81.2080.40 StarVLA4B88.2088.30 π 0 3B65.9258.40 π 0.5 3B82.7476.76 X-VLA0.9B72.8872.84 Motus8B88.6687.02 GigaWorld-Policy 5B86.3685.04 Being-H0.73B90.2089.60 Ours0.5B90.4889.04 Table 1: Average Success rates (%) over 50 RoboTwin 2.0 tasks. MethodParams Spatial Object Goal Long AVG OpenVLA7B84.788.479.2 53.7 76.5 OpenVLA-OFT 7B97.698.497.9 94.5 97.1 StarVLA4B97.898.696.2 93.8 96.6 π 0 3B96.898.895.8 85.2 94.1 GR00T-N1.63.3B97.798.597.5 94.4 97.0 JEPA-VLA–97.298.095.6 94.8 96.4 SmolVLA2B93.094.091.0 77.0 88.8 EVO-10.8B92.797.796.3 92.3 94.8 Ours0.5B98.098.897.2 94.2 97.1 Table 2: Success rates (%) on the four LIBERO suites (Spatial, Object, Goal, and Long). 4 Experiments 4.1 Implementation Details LiLa-WAM builds on a frozen DINOv3-ViT-L/16 encoder [43], using features from its 17th and 21st blocks. The visual features are processed by a 64-query adapter with 4 layers, followed by the Foresight-Aware Action Expert, a 12-layer DiT with 8 attention heads and a feature dimension of 768, and a 4-layer decoder. The policy takes a 16-dimensional proprioceptive state as input and predicts a 14-dimensional action sequence with a chunk length of 32, of which the first 16 steps are executed. The camera resolution is320× 240. The future horizon∆is set to the action chunk length, i.e., ∆ = 32. The model contains 0.5B parameters, of which 0.2B are trainable and 0.3B are frozen. We train with AdamW [47] (β 1 = 0.9,β 2 = 0.99, weight decay 0.01), a batch size of 128, gradient clipping with a maximum norm of 1.0, andλ f = 0.5. During flow-matching training, the timestep is sampled from a logit-normal distribution. We adopt a stage-wise cosine learning-rate schedule, using a peak learning rate of2× 10 −4 in the early stage and5× 10 −5 in the later stage. All training fits within a single 24 GB GPU budget: joint training over the 50 RoboTwin 2.0 tasks takes about 110 GPU hours on an NVIDIA RTX 5090. At deployment, the VTT corresponding to the target task is selected, and action chunks are generated by integrating the learned velocity field with the Euler method using 10 ODE steps. Inference takes 85 ms on an NVIDIA RTX 4090. 4.2 Simulation Environments RoboTwin 2.0. RoboTwin 2.0 [48] serves as our primary benchmark, covering 50 manipulation tasks. For each task, we collect 50 clean demonstrations and 500 randomized demonstrations with domain randomization, and train a single model jointly on the data of all tasks. The policy receives a single third-person RGB view resized to320× 240, and predicts action chunks of size 32. Each task is evaluated over 50 rollouts. For ablation studies, we additionally construct a subset of 10 representative tasks spanning different types and difficulty levels: adjust_bottle, grab_roller, hanging_mug, move_stapler_pad, open_microwave, press_stapler, scan_object, stack_blocks_two, stamp_seal, and turn_switch. LIBERO.We further evaluate on LIBERO [49] to compare with lightweight VLA baselines. The policy takes both a third-person view and a wrist-mounted view resized to256× 256, and predicts action chunks of size 12. A separate policy is trained for each of the four suites, each containing 10 tasks (40 tasks in total), and each task is evaluated over 50 rollouts. Results on RoboTwin 2.0.Table 1 compares LiLa-WAM with recent state-of-the-art models [16,19,31,32,50–53] on the 50 tasks of RoboTwin 2.0. Using a frozen DINOv3 encoder, LiLa-WAM is trained on a single GPU and remains competitive with larger manipulation policies. It surpasses the 8B Motus and 5B GigaWorld-Policy by 1.8 and 4.1 points, respectively, while using 16×and 10×fewer parameters. In the clean setting, it slightly exceeds Being-H0.7 by 0.3 points. Results on LIBERO. Table 2 compares LiLa-WAM with recent state-of-the-art models [17,19,24,25,53–56] on LIBERO [49]. At 0.5B parameters, LiLa-WAM attains a 97.1% average success rate, matching the baseline OpenVLA- OFT with roughly 14×fewer parameters and outperformingπ 0 by 3.0 points. Among lightweight models, it surpasses 6 EVO-1 and SmolVLA by 2.3 and 8.3 points. These results suggest that foresight-aware manipulation can be realized with a lightweight, end-to-end trainable architecture under a single-GPU budget. 4.3 Ablation Study Foresight supervision, task conditioning, and backbone. Table 3 presents three comparisons. First, removing the foresight objectiveL f collapses the model to a reactive VLA and drops the success rate from 70.0% to 54.4%, indicating that foresight supervision is an important contributor to performance. Second, replacing the VTT with CLIP-encoded language instructions degrades performance by 8.6 points, showing that the VTT offers a viable lightweight form of task specification. Third, a Qwen3VL-2B backbone reaches 61.0% despite being over 4×larger, showing that a larger pretrained backbone alone does not guarantee better control performance under the same training budget. BackboneForesightTask Cond.ParamsAVG DINOv3 ×VTT0.5B54.4 DINOv3✓CLIP-lang0.5B61.4 Qwen3VL✓Language2.2B61.0 DINOv3✓VTT0.5B70.0 Table 3: Ablation on foresight supervision, task conditioning, and visual backbone. Average success rate over the 10 RoboTwin 2.0 tasks. The last row is our full model. Foresight supervision and the VTT each contribute substantial gains, and the lightweight DINOv3 backbone outperforms a larger VLM backbone. Number of adapter queries.Table 4 reports the effect of the adapter query countN q , withN q = 64performing best (70.0%). Smaller budgets (8 or 32) likely lose spatial detail needed for fine-grained manipulation, whereas 96 tokens slightly degrades performance while increasing memory, suggesting diminishing returns beyond a sufficient latent size. Training memory grows steadily from 14.7 GB to 21.3 GB asN q increases. We thus adoptN q = 64, combining the best performance with a comfortable margin under the 24 GB budget. N q (Adapter)8326496 AVG66.265.270.067.6 Training Memory (GB)14.716.018.021.3 Table 4: Effect of adapter query count N q on success rate and training memory on 10 RoboTwin 2.0 tasks. Choice of DINOv3 feature layers.Table 5 examines which encoder layers provide the most control-relevant features. Performance improves steadily as features are drawn from earlier blocks, rising from 48.0% with the final block to 67.2% with the 17th. A likely explanation is that later blocks emphasize global semantics at the expense of spatial detail, whereas mid-level blocks better preserve the object locations and local geometry that fine-grained manipulation requires. Combining the 17th and 21st blocks further improves the success rate to 70.0%, indicating that the two levels offer complementary cues. Feature Layers2423211717 & 21 Avg48.053.257.467.270.0 Table 5: Effect of DINOv3 feature layers. Average success rate over the 10 RoboTwin 2.0 tasks when aggregating different feature layers of the encoder. 4.4 What Do the VTT and Action Tokens Attend To? Attention analysis.Figure 4 compares the attention distributions of the VTT and the action tokens on the input image. Interestingly, the two token types exhibit a division of labor: the VTT focuses more on the objects that the task acts upon (e.g., the fan and its target pad, the block and the hammer), while the action tokens attend more to the arm and gripper, with responses spreading over the manipulation region. This pattern matches the intuition that task specification should capture what to manipulate, whereas action generation track how the manipulator moves, suggesting that the VTT provides task-relevant cues. 7 P lace F an VTTAction Tokens B eat B lock H ammer Step 1 Step 2 Step 2 Step 1 Observation Open M icrowave Figure 4: Attention maps of the VTT and action tokens. Layer 1 - 6 VisualTransition Token (VTT) Attention Map Action Tokens Attention Map Layer 7 - 12 Layer 1 - 6 Layer 7 - 12 Figure 5: Per-layer attention maps of the VTT and action tokens. Compared with the action tokens, the VTT attends more to the placement area, whereas the action tokens attend more to the object being manipulated (red boxes). Per-layer attention analysis. Figure 5 visualizes, for each of theL=12DiT blocks of the action expert, how the VTT and the action tokens distribute their attention over the image. The visualization is conducted on the place fan task from RoboTwin 2.0, which requires grasping a small fan and placing it onto a marked pad. We extract the attention weights from each token type to the visual tokens in every block; since inference involves multiple denoising steps, the action-token maps are averaged over all denoising steps. Two observations emerge. First, the VTT activates on both the fan to be grasped and the pad marking its destination. Second, the action tokens concentrate more narrowly on the fan currently being manipulated. Together, the two token types exhibit a complementary division of labor: the VTT provides task-level context, while the action tokens focus on the object at hand, indicating that the VTT provides task-level guidance. 4.5 Action Conditioning in the Action Expert Action perturbation probeTo examine whether the predicted future-state tokens encode action-conditioned dynam- ics, we intervene on the action tokens along the flow-matching trajectory. At each denoising timesteps, we either inject additional Gaussian noise into the action tokens or replace the whole action chunk with an action sequence sampled from another trajectory. Since the future-state tokens are compressed by the adapter, we use the decoder-reconstructed tokens as the readout and measure the cosine similarity between predictions obtained with the original and perturbed actions. A similarity of 1 indicates that the prediction is insensitive to the intervention. As shown in Figure 6, both perturbations have little effect during the early denoising stage, where the action tokens are still dominated by noise. In contrast, replacing the action chunk leads to a clear similarity drop afters≈ 0.5, reaching approximately0.85at s = 1, whereas adding noise only causes a marginal decrease. This suggests that the predicted future features are not 8 determined solely by the visual context, but encode action-conditioned future dynamics. The stronger sensitivity at late denoising stages is expected: as denoising progresses, the action tokens become semantically specific and the model gradually commits to a particular action trajectory, so replacing them introduces an inconsistent future. We further observe that the attention from visual-token positions to action tokens remains low across different denoising stages. This indicates that the late-stage sensitivity is not caused by a small number of high-attention connections, but rather by distributed information mixing through self-attention. 0.00.20.40.60.81.0 t (0 = pure noise, 1 = clean action) 0.75 0.80 0.85 0.90 0.95 1.00 Decoder output cosine similarity (baseline vs perturbed) (1 = insensitive) replace noise Figure 6: Action perturbation probe of future feature prediction. We perturb the action tokens at different denoising timesteps either by adding Gaussian noise or by replacing the entire action chunk with an action sequence sampled from another trajectory. The decoder-reconstructed future-feature tokens are used as the readout, and we report the cosine similarity between predictions obtained with the original and perturbed actions. Attention mask analysis.Table 6 reports the effect of suppressing the attention from visual-context tokens to action tokens on 10 RoboTwin 2.0 tasks. Under the default bidirectional self-attention, visual-context tokens can use action tokens as keys and values, allowing future feature prediction to obtain action-related information. When this attention pathway is suppressed, the average success rate decreases from 70.0% to 64.8%. This suggests that the action-to-visual information flow is important for dynamics-aware prediction, providing additional evidence that the predicted future features are action-conditioned. Notably, this degradation occurs even though the corresponding attention weights remain relatively low across denoising steps, indicating that useful action information may be transmitted through sparse or distributed attention patterns rather than a small number of dominant connections. Attention SettingMasked Query–KVAvg. Success (%) Bidirectional attention–70.0 Visual-to-action attention maskQ:Z v , K/V:Z a 64.8 Table 6: Effect of suppressing the attention from visual-context tokens to action tokens on 10 RoboTwin 2.0 tasks.Z v denotes the compressed visual-context tokens, and Z a denotes the action tokens. Attention dynamics across denoising steps.The Foresight-Aware Action Expert applies bidirectional self-attention over the full token sequence without any attention mask. All token groups can thus freely exchange information. Figure 7 tracks the mean attention between token groups across the denoising steps, averaged over all DiT blocks. Three trends stand out. First, the VTT consistently assigns most of its attention to the reasoning tokens throughout denoising. This suggests that the task representation maintains interaction with the visual context throughout denoising. Second, the action tokens attend strongly to the reasoning tokens at early steps. This reliance gradually decreases as denoising proceeds, which may reflect a transition from context gathering to action refinement. Third, attention from the reasoning tokens to the action tokens remains low throughout denoising. 4.6 Real-Robot Experiments Setup. As shown in Figure 8, we conduct real-robot experiments on an Agilex Piper 6-DoF robotic arm, using an Intel RealSense D435 camera as the third-person view. The policy receives RGB observations resized to256× 256 9 12345678910 Denoising Step 0.1 0.2 0.3 0.4 0.5 Mean Attention Action Reasoning Action VTT VTT Reasoning VTT Action Reasoning Action Figure 7: Attention dynamics across denoising steps on the press stapler task from RoboTwin 2.0. Mean attention between token groups at each of the 10 flow-matching denoising steps, averaged over all DiT blocks. and outputs absolute joint-angle commands at 30 Hz. The four tasks in Table 7 are evaluated with randomized object placements to test robustness to variations in the initial scene configuration. For each task, we collect 60 demonstrations and train both variants on the same data. Each method is evaluated over 50 rollouts per task, and we report the success rate. We compare LiLa-WAM with its variant without the foresight lossL f , which isolates the contribution of future-state supervision on the physical robot. More setup details are provided in the supplementary material. Put the banana on the pink platePut the basketball in the hoop Put the red block on the boxPush the green button Figure 8: Execution process of the four real-robot tasks. Results. Table 7 reports the real-robot results. LiLa-WAM achieves higher success rates than the ablated variant on all four tasks, improving the average success rate from 74.0% to 82.0%. The gains range from 4 to 14 points, with the largest improvement observed on push the green button. Put the basketball in the hoop remains the most challenging task for both variants. These results indicate that foresight supervision also benefits real-world robotic control, although performance on more difficult tasks still leaves room for improvement. MethodButtonBlockBananaBasketballAvg w/oL f 7284924874.0 LiLa-WAM8690965682.0 Table 7: Real-robot success rates (%). The four tasks are push the green button (Button), put the red block on the box (Block), put the banana on the pink plate (Banana), and put the basketball in the hoop (Basketball). 10 5 Conclusion In this paper, we propose LiLa-WAM, a lightweight world-action model that reasons about the future in a compact latent space jointly shaped by future-state prediction and action generation, and can be trained end-to-end on a single GPU. We further introduce the Visual Transition Token, a language-free task representation that encodes each task as a direction in visual feature space. Experiments on RoboTwin 2.0, LIBERO, and real-robot tasks, together with ablation and visualization analyses, support the effectiveness of the proposed model. References [1]Yucheng Hu, Yanjiang Guo, Pengchao Wang, Xiaoyu Chen, Yen-Jen Wang, Jianke Zhang, Koushil Sreenath, Chaochao Lu, and Jianyu Chen. Video prediction policy: A generalist robot policy with predictive visual representations. arXiv preprint arXiv:2412.14803, 2024. [2]Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in neural information processing systems, pages 9156–9172, 2023. [3] Yang Tian, Sizhe Yang, Jia Zeng, Ping Wang, Dahua Lin, Hao Dong, and Jiangmiao Pang. Predictive inverse dynamics models are scalable learners for robotic manipulation. In International Conference on Learning Representations, pages 92033–92052, 2025. [4]Shalfun Li, Victor Yao, Charles Yang, Truth Qu, Regis Cheng, Ryan Yu, Howard Lu, Newton Von, Vincent Chen, Yohann Tang, et al. Wall-wm: Carving world action modeling at the event joints. arXiv preprint arXiv:2606.01955, 2026. [5] Jisong Cai, Long Ling, Shiwei Chu, Zhongshan Liu, Jiayue Kang, Zhixuan Liang, Wenjie Xu, Yinan Mao, Weinan Zhang, Xiaokang Yang, et al. Aha-wam: Asynchronous horizon-adaptive world-action modeling with observation-guided context routing. arXiv preprint arXiv:2606.09811, 2026. [6]Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, Feihong Liu, Chenchen Yang, Yikang Zhou, Zhaoye Fei, Jingjing Gong, Jinlan Fu, et al. World action models: The next frontier in embodied ai. arXiv preprint arXiv:2605.12090, 2026. [7]Yuyang Zhang, Wenyao Zhang, Zekun Qi, He Zhang, Haitao Lin, Jingbo Zhang, Yao Mu, Xiaokang Yang, Wenjun Zeng, and Xin Jin. Imagewam: Do world action models really need video generation, or just image editing? arXiv preprint arXiv:2606.19531, 2026. [8] Physical Intelligence, Bo Ai, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Greg Balke, Kevin Black, George Bokinsky, Shihao Cao, Thomas Charbonnier, et al. pi0.7: a steerable generalist robotic foundation model with emergent capabilities. arXiv preprint arXiv:2604.15483, 2026. [9]Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming-Yu Liu, Chelsea Finn, et al. Cosmos policy: Fine-tuning video models for visuomotor control and planning. arXiv preprint arXiv:2601.16163, 2026. [10]Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control. arXiv preprint arXiv:2601.21998, 2026. [11]Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026. [12]Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Se June Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. In International Conference on Learning Representations, pages 28213–28239, 2025. [13]Ruijie Zheng, Jing Wang, Scott Reed, Johan Bjorck, Yu Fang, Fengyuan Hu, Joel Jang, Kaushil Kundalia, Zongyu Lin, Loic Magne, et al. Flare: Robot learning with implicit world modeling. arXiv preprint arXiv:2505.15659, 2025. [14] Han Zhao, Jingbo Wang, Wenxuan Song, Shuai Chen, Yang Liu, Yan Wang, Haoang Li, and Donglin Wang. Frappe: Infusing world modeling into generalist policies via multiple future representation alignment. arXiv preprint arXiv:2602.17259, 2026. [15] Yue Su, Sijin Chen, Haixin Shi, Mingyu Liu, Zhengshen Zhang, Ningyuan Huang, Weiheng Zhong, Zhengbang Zhu, Yuxiao Liu, and Xihui Liu. World guidance: World modeling in condition space for action generation. arXiv preprint arXiv:2602.22010, 2026. 11 [16]Hao Luo, Wanpeng Zhang, Yicheng Feng, Sipeng Zheng, Haiweng Xu, Chaoyi Xu, Ziheng Xi, Yuhui Fu, and Zongqing Lu. Being-h0.7: A latent world-action model from egocentric videos. arXiv preprint arXiv:2605.00078, 2026. [17]Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. [18] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. [19]Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. [20] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, (10-11):1684–1704, 2025. [21] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. [22] Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024. [23]Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Zhibin Tang, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. IEEE Robotics and Automation Letters, 2025. [24] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. [25]Tao Lin, Yilei Zhong, Yuxin Du, Jingjing Zhang, Jiting Liu, Yinxinyu Chen, Encheng Gu, Ziyan Liu, Hongyi Cai, Yanwen Zou, et al. Evo-1: Lightweight vision-language-action model with preserved semantic alignment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13397–13406, 2026. [26]Jonas Pai, Liam Achenbach, Victoriano Montesinos, Benedek Forrai, Oier Mees, and Elvis Nava. mimic-video: Video-action models for generalizable robot control beyond vlas. arXiv preprint arXiv:2512.15692, 2025. [27] Yao Feng, Hengkai Tan, Xinyi Mao, Chendong Xiang, Guodong Liu, Shuhe Huang, Hang Su, and Jun Zhu. Vidar: Embodied video diffusion model for generalist manipulation. arXiv preprint arXiv:2507.12898, 2025. [28]Yue Liao, Pengfei Zhou, Siyuan Huang, Donglin Yang, Shengcong Chen, Yuxin Jiang, Yue Hu, Jingbin Cai, Si Liu, Jianlan Luo, et al. Genie envisioner: A unified world foundation platform for robotic manipulation. arXiv preprint arXiv:2508.05635, 2025. [29] Jun Cen, Chaohui Yu, Hangjie Yuan, Yuming Jiang, Siteng Huang, Jiayan Guo, Xin Li, Yibing Song, Hao Luo, Fan Wang, et al. Worldvla: Towards autoregressive action world model. arXiv preprint arXiv:2506.21539, 2025. [30]Shuang Li, Yihuai Gao, Dorsa Sadigh, and Shuran Song. Unified video action model. arXiv preprint arXiv:2503.00200, 2025. [31]Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, et al. Motus: A unified latent action world model. arXiv preprint arXiv:2512.13030, 2025. [32]Angen Ye, Boyuan Wang, Chaojun Ni, Guan Huang, Guosheng Zhao, Hao Li, Hengtao Li, Jie Li, Jindi Lv, Jingyu Liu, et al. Gigaworld-policy: An efficient action-centered world–action model. arXiv preprint arXiv:2603.17240, 2026. [33]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. [34]Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, et al. Lawam: Latent world action models for efficient dynamics-aware robot policies. arXiv preprint arXiv:2606.15768, 2026. 12 [35]Pengfei Zhou, Liliang Chen, Shengcong Chen, Di Chen, Wenzhi Zhao, Rongjun Jin, Guanghui Ren, and Jianlan Luo. Act2goal: From world model to general goal-conditioned policy. arXiv preprint arXiv:2512.23541, 2025. [36]Haonan Chen, Jingxiang Guo, Bangjun Wang, Tianrui Zhang, Xuchuan Huang, Boren Zheng, Yiwen Hou, Chenrui Tie, Jiajun Deng, and Lin Shao. Goal-vla: Image-generative vlms as object-centric world models empowering zero-shot robot manipulation. arXiv preprint arXiv:2506.23919, 2025. [37]Hongyin Zhang, Pengxiang Ding, Shangke Lyu, Ying Peng, and Donglin Wang. Gevrm: Goal-expressive video generation model for robust visual manipulation. arXiv preprint arXiv:2502.09268, 2025. [38]Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. In 8th Annual Conference on Robot Learning, 2024. [39] Zhuoling Li, Liangliang Ren, Jinrong Yang, Yong Zhao, Xiaoyang Wu, Zhenhua Xu, Xiang Bai, and Hengshuang Zhao. Vip: Vision instructed pre-training for robotic manipulation. arXiv preprint arXiv:2410.07169, 2024. [40]Cunxin Fan, Xiaosong Jia, Yihang Sun, Yixiao Wang, Jianglan Wei, Ziyang Gong, Xiangyu Zhao, Masayoshi Tomizuka, Xue Yang, Junchi Yan, et al. Interleave-vla: Enhancing robot manipulation with interleaved image-text instructions. arXiv preprint arXiv:2505.02152, 2025. [41]Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631, 2025. [42]Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. [43]Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025. [44]Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10371–10381, 2024. [45]Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017. [46]Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. [47]Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. International Conference on Learning Representations, 2019. [48]Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. arXiv preprint arXiv:2506.18088, 2025. [49]Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, pages 44776–44791, 2023. [50] Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, et al. X-vla: Soft-prompted transformer as scalable cross-embodiment vision-language- action model. arXiv preprint arXiv:2510.10274, 2025. [51] Yandan Yang, Shuang Zeng, Tong Lin, Xinyuan Chang, Dekang Qi, Junjin Xiao, Haoyun Liu, Ronghan Chen, Yuzhi Chen, Dongjie Huo, et al. Abot-m0: Vla foundation model for robotic manipulation with action manifold learning. arXiv preprint arXiv:2602.11236, 2026. [52]Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.π 0.5 : a Vision-Language-Action Model with Open-World Generalization. arXiv preprint arXiv:2504.16054, 2025. [53] StarVLA Community. Starvla: A lego-like codebase for vision-language-action model developing. arXiv preprint arXiv:2604.05014, 2026. [54]Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. arXiv preprint arXiv:2502.19645, 2025. 13 [55]Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734, 2025. [56]Shangchen Miao, Ningya Feng, Jialong Wu, Ye Lin, Xu He, Dong Li, and Mingsheng Long. Jepa-vla: Video predictive embedding is needed for vla models. arXiv preprint arXiv:2602.11832, 2026. 14 A Real-Robot Experiment Details Hardware platform. Real-robot experiments are conducted on an AgileX PiPer robotic arm, a 6-DoF manipulator equipped with a parallel gripper. The action space consists of absolute joint positions, and the gripper is controlled with continuous values. A RealSense D435 camera provides RGB observations from a third-person viewpoint, mounted at a head-like position overlooking the workspace. The policy predicts action chunks of size 32, and the robot operates at a control frequency of 30 Hz. Asynchronous inference. To prevent execution pauses caused by inference latency, we adopt an asynchronous inference scheme with multi-threading. A dedicated thread maintains a queue of actions to be executed. When the number of remaining actions in the queue drops to 8, model inference is triggered in a separate thread; upon completion, the queue is refreshed with the newly predicted action chunk. In this way, the robot always has pending actions to execute while the model is inferring, which eliminates pauses and ensures the continuity of motion. All real-robot experiments, including the baseline without future-state supervision, use identical data preprocessing and postprocessing pipelines. This controlled setting helps isolate the effect of future-state supervision from differences in data handling. fold_towel hanging_mug push_the_green_button put_the_apple_in_the_blue_bowl put_the_banana_on_the_pink_plate put_the_basketball_in_the_hoop put_the_black_chain_in_the_box put_the_red_block_on_the_box put_the_red_block_in_the_purple_cup put_the_red_block_on_the_white_bowl put_the_toothbrush_in_the_yellow_cup put_the_white_spoon_in_the_purple_cup stack_the_red_block_on_the_wooden_block stack_the_yellow_block_on_the_wooden_block Figure 9: t-SNE visualization of per-episode visual transition embeddings on the real-robot demonstration set, where each point corresponds to one episode. open_the_middle_drawer_of_the_cabinet open_the_top_drawer_and_put_the_bowl_inside push_the_plate_to_the_front_of_the_stove put_the_bowl_on_the_plate put_the_bowl_on_the_stove put_the_bowl_on_top_of_the_cabinet put_the_cream_cheese_in_the_bowl put_the_wine_bottle_on_the_rack put_the_wine_bottle_on_top_of_the_cabinet turn_on_the_stove Figure 10: t-SNE visualization of per-episode visual transition embeddings on LIBERO-goal. Each point represents the visual transition embedding of a single episode. Semantically related tasks that share objects or scenes (e.g., put the wine bottle on top of the cabinet and put the wine bottle on the rack) tend to have closer clusters, suggesting that the visual transition embeddings capture both task-relevant transition information and task similarity. 15 π 0.5 MotusX-VLAGigaWorld-Policy LiLa-WAM TaskClean Rand. Clean Rand. Clean Rand. CleanRand.Clean Rand. Adjust Bottle1009989931009910010010098 Beat Block Hammer96939588928886869690 Blocks Ranking RGB9285999783839296100100 Blocks Ranking Size49267563677444489288 Click Alarmclock988910010099991001009492 Click Bell99661001001001001001009090 Dump Bin Bigbin929795917977921009694 Grab Roller100100100100100100100100100100 Handover Block66578673733780809690 Handover Mic9897786300727210098 Hanging Mug18173838232716125644 Lift Pot9685969999100989810096 Move Can Pot51553474898676789890 Move Pillbottle Pad84619396737190909894 Move Playingcard Away96841009693987872100100 Move Stapler Pad56428385787392826676 Open Laptop9096959193100969894100 Open Microwave34779591797174666658 Pick Diverse Bottles81719091583682708488 Pick Dual Bottles9363969047368686100100 Place A2B Left87828879484994886878 Place A2B Right87849187363690927676 Place Bread Basket77649194817182829498 Place Bread Skillet85668683776794909288 Place Burger Fries94879898949498969696 Place Can Basket62628176495278748072 Place Cans Plasticbox94849894979810010098100 Place Container Plate99959899979598969698 Place Dual Shoes75759387798896846054 Place Empty Cup100999998100989090100100 Place Fan87859187807592949690 Place Mouse Pad60396668707088908876 Place Object Basket80768187443990928890 Place Object Scale86808885527488809688 Place Object Stand918598978688100989294 Place Phone Stand81818786888782729294 Place Shoe929399979695989610098 Press Stapler878393989298969610098 Put Bottles Dustbin84798179747772709294 Put Object Cabinet80798871464874749292 Rotate QRcode89878973343390848884 Scan Object72656766143660649490 Shake Bottle99971009799100100100100100 Shake Bottle Horizontally99991009810010010098100100 Stack Blocks Three9176919561070788678 Stack Blocks Two971001009892871009410098 Stack Bowls Three77717987768670728882 Stack Bowls Two959698989693969298100 Stamp Seal79559392768296987278 Turn Switch62548478406182847682 Average82.74 76.76 88.66 87.02 72.88 72.84 86.3685.0490.48 89.04 Table 8: Per-task success rates (%) on 50 RoboTwin 2.0 tasks under clean and randomized settings. 16