{"id":"08db9ff3-875f-4a84-b107-b448903cf4be","arxiv_id":"2609.20709","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"MoWAM replaces costly future video generation with explicit future motion prediction in world action models, achieving strong OOD robustness with low inference cost.","lead":"This paper proposes MoWAM, a robot control model that predicts future robot motion instead of generating full future videos, reducing inference cost while maintaining robustness to novel situations. The method achieves strong performance on standard benchmarks and real-world tasks, and offers a practical way to scale test-time computation by evaluating multiple candidate motion sequences.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Real-world keypoint acquisition method is unspecified; the claimed real-world applicability depends on an unstated assumption about keypoint availability and accuracy.","rationale":"The reader's weakest_assumption focuses on the sufficiency of motion as an abstraction of future dynamics. I agree that this is a concern, but I identify a more concrete and immediate missing specification: how keypoints are acquired in real-world settings. This is a prerequisite for the reader's assumption—if keypoints cannot be reliably obtained, the motion abstraction is irrelevant. The reader's CONDITIONAL verdict already invites further validation; my test adds a specific, actionable condition. Therefore, the verdict remains CONDITIONAL (unchanged).","tokens_in":12911,"tokens_out":2475,"duration_ms":24419,"concrete_test":"Require the authors to (a) specify the exact method for obtaining fingertip and palm keypoints in the real-world experiments (e.g., forward kinematics with camera projection, or a pretrained keypoint estimator), and (b) report keypoint detection accuracy (e.g., mean pixel error) on a held-out set of real-world images. Then, re-run the three real-world tasks with keypoints perturbed by adding Gaussian noise (σ = 5 pixels) to test sensitivity. If success rates drop by more than 10 percentage points, the method's practical robustness is limited and should be stated as a limitation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's motion representation (Section III-B) uses projected fingertip and palm keypoint positions in image space. In simulation (LIBERO), these are readily available from the simulator state. For real-world experiments (Table II, Section IV-A1), the paper does not specify how these keypoints are obtained. Without describing the detection or projection method, the reproducibility and generalizability of the real-world results are unclear. If keypoints are obtained via forward kinematics and camera projection, this assumes calibrated cameras and accurate robot pose; if a learned detector is used, its accuracy and robustness to occlusions or lighting changes are unknown. The entire method—motion prediction, action–motion conditioning, and verifier—depends on accurate keypoints at both training and inference. This missing detail is load-bearing because the real-world evaluation is a key part of the central claim (\"higher average real-world success than representative WAM baselines\"). Without it, the reader cannot assess whether the method's success is contingent on an unstated, potentially fragile pipeline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes MoWAM, a World Action Model that replaces computationally expensive future video generation at inference with explicit prediction of structured end-effector motion (fingertip and palm keypoint geometry changes). A Mixture-of-Transformer architecture is trained with future video modeling but at inference only predicts motion and action, using a motion-aware verifier for inference-time candidate scaling. Experiments on LIBERO (98.9% avg.), LIBERO-Plus (81.43% avg., outperforming Fast-WAM by 11 points), and three real-world tasks (80% avg. success) show strong in-distribution and OOD performance with lower latency than full video-generation baselines.","tokens_in":13190,"tokens_out":1655,"duration_ms":16008,"significance":"If the results hold, MoWAM offers a principled and practical trade-off between explicit future reasoning and inference cost for World Action Models. The compact motion representation makes inference-time candidate exploration feasible and yields consistent gains. The method is architecture-agnostic and could generalize to other domains where robot motion is a sufficient abstraction of future dynamics. The real-world validation and ablation studies strengthen the claim that explicit structured motion captures task-relevant future information without dense visual generation.","major_comments":[{"comment":"The method depends critically on obtaining projected fingertip and palm keypoints (p^j_{t,l}, p^j_{t,r}, p^j_{t,p}) in image space. For the LIBERO simulation these are trivially available from the simulator state. For the real-world experiments (Table II, Section IV-A1), the paper does not specify how these keypoints are acquired—whether through forward kinematics with camera calibration, a learned keypoint detector, or manual annotation. This omission is load-bearing because the entire motion representation, motion prediction, and verifier rely on accurate keypoints at both training and inference. Without this detail, readers cannot assess reproducibility, robustness to occlusion or calibration error, or whether the method's real-world success depends on an unstated and potentially fragile pipeline. The authors should describe the keypoint acquisition method, its accuracy, and failure modes.","section":"Section IV-A1 (Real-World Evaluation) and Section III-B"},{"comment":"The verifier is trained by ranking demonstrated motions (+1) against perturbed motions (-1), where perturbations include 'reducing motion changes, adding a smooth spatial shift, or cutting off the later part of the motion.' The choice of perturbations seems reasonable, but the paper does not report the verifier's accuracy or agreement with task success. If the verifier systematically favors motions that match the training distribution rather than those that lead to task completion, it could introduce a circularity that inflates scaling results. The authors should add a quantitative analysis of the verifier's predictive performance on held-out demonstrations or perturbed trajectories.","section":"Section III-D (Verifier Training)"}],"minor_comments":[{"comment":"Equation (4) defines m_{dyn,t} in R^6, and Equation (5) defines m_{str,t} in R^4. However, the paper later states 'm = (m_dyn, m_str)' without specifying the full dimension when aggregated over H time steps and multiple camera views. Clarify the total dimensionality and how multiple camera views are fused.","section":"Section III-B"},{"comment":"The arrow labeled 'Training Stage Only' is ambiguous: does it indicate that the green path (future frames, keypoints projection, motion abstraction) is used only during training? The caption should explicitly state that video generation and keypoint projection occur only in training.","section":"Figure 2"},{"comment":"The row labels 'w/o Dynamics Loss' and 'w/o Structure Loss' refer to two versions of motion supervision ablation. The parenthesized percentages like '(3.43↑)' are not defined; make clear that these indicate the absolute increase over the previous ablated baseline.","section":"Table IV"},{"comment":"The method name appears inconsistently as 'MoW AM' (with space) and 'MoWAM' (without). Unify throughout the manuscript.","section":"General"},{"comment":"The conclusion restates the contributions but does not discuss limitations. Consider adding a brief limitations paragraph (e.g., reliance on keypoint accuracy, assumption that robot motion is sufficient abstraction for all tasks).","section":"Section V"}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Jiayu and colleagues propose a clean idea: instead of generating full future videos at inference (costly) or dropping future prediction entirely (losing robustness), represent the future as structured end-effector motion — fingertip/palm keypoint displacements and geometry changes. This is new and sensible. The Mixture-of-Transformer architecture that learns video dynamics during training but only predicts motion+action at inference is well engineered. The experiments are the strongest part: extensive results on LIBERO (98.9%), LIBERO-Plus (81.43%), three real-world tasks (80%), plus thorough ablations that isolate contributions of motion supervision, video training, and motion conditioning. The inference-time scaling demonstration, though limited to one task, shows the practical benefit of the compact representation.\n\nTwo soft spots stand out, one of them load-bearing. First, the real-world keypoint acquisition is not described anywhere. In simulation, keypoints come from the simulator state; for real-world experiments, the paper does not say whether they are obtained via forward kinematics + calibrated camera, a learned keypoint detector, or some other method. The entire method depends on accurate keypoints at both training and inference, so this omission undermines reproducibility. The stress-test note is correct. Second, the verifier training is described in only a paragraph — the perturbation scheme and margin are given, but details like data volume, training procedure, and architecture are missing. These are secondary but worth addressing.\n\nThe motion-sufficiency assumption (that keypoint changes capture all task-relevant future dynamics) is a limitation the authors acknowledge indirectly by discarding visual details. They do not test on tasks involving deformable objects or fine-grained physical interaction, so the scope is narrower than the title might imply. That is a minor concern, not a flaw, as the paper claims strong results on the tasks it evaluates.\n\nOverall, this is a solid paper for robotics researchers working on efficient world models. It deserves a serious referee. With the keypoint acquisition clarified and verifier details filled in, it should be published.","headline":"MoWAM makes a genuine contribution by replacing dense future video generation with structured motion prediction, but the unspecified real-world keypoint acquisition method is a load-bearing gap that must be addressed before the experiments can be fully assessed.","tokens_in":13718,"tokens_out":1631,"duration_ms":16032},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Replacing future video generation with explicit motion prediction yields efficient and robust robot policies.","keywords":["robot manipulation","world action model","future motion prediction","inference-time scaling","mixture of transformers","structured representation","distribution robustness","efficient robot policy"],"falsifier":"Test MoWAM on a manipulation task involving a deformable object (e.g., folding a cloth or squeezing a sponge) where the end-effector geometry stays nearly constant but the object shape changes; if success rate drops sharply compared to a full-video WAM, the motion abstraction is insufficient for such tasks.","tokens_in":14492,"feed_emoji":"🤖","tokens_out":3124,"duration_ms":29379,"temperature":0.7,"pith_summary":"World Action Models (WAMs) improve robot control by predicting how the scene will evolve, but generating full future videos at inference is computationally expensive. MoWAM replaces that dense video generation with explicit prediction of structured robot motion, captured by the changing geometry of fingertip and palm keypoints. This compact motion representation retains the benefits of explicit future reasoning while cutting latency by an order of magnitude. It also enables efficient inference-time scaling: sampling multiple action-motion candidates and selecting the best one via a learned verifier. Experiments show MoWAM matching or exceeding prior WAMs on standard benchmarks, with particular gains under distribution shifts.","feed_headline":"Robot motion prediction replaces costly video generation for smarter policies","feed_subtitle":"","key_machinery":"Structured end-effector motion, defined by temporal changes in gripper center, fingertip span, and palm offset from two camera views, serves as a compact explicit future representation. The model has a dual-branch Transformer: a Video Transformer learns future visual dynamics during training (discared at inference) and provides observation keys/values; the Action-Motion Transformer jointly denoises action and motion tokens conditioned on those keys/values, with flow matching for motion distribution modeling.","core_discovery":"For robot manipulation, the control-relevant information in a predicted future scene is largely captured by how the end effector is expected to move, not by the full visual appearance. MoWAM uses this insight to define a structured motion representation based on geometric changes (displacement and relative geometry) of fingertip and palm keypoints. A ixture-of-Transformer architecture jointly learns future video dynamics during training and predicts both action and motion independently of generated future frames, allowing video generation to be removed at inference while still providing an explicit future. A motion-aware verifier then scores multiple motion-action pairs for inference-time scaling, demonstrating that performance improves with more candidates at minimal additional cost.","pith_inferences":["MoWAM's assumption that all future control-relevant information is contained in end-effector geometry may break down for tasks requiring deformation of objects (e.g., folding cloth) or those where object positions change non-rigidly; in such cases, the motion abstraction would be insufficient.","The motion verifier is trained with hand-crafed negtive examples; a more principled approach (e.g., leveraging the model's own confidence or contrastive learning between predicted and actual motion) could improve its generalization to unseen distribution shifts.","The method still requires a pretained video model (Wan) with substantial paramers (5B), which may not be feasible for low-resource settings; distilling motion prediction into a lighter architecture is a natural extension not explored in the paper."],"forward_implications":["WAMs can achieve competitive in-distribution and out-of-distribution robustness without generating any future video at inference, reducing latency from hundreds of milliseconds to under 300 ms.","The compact motion representation allows practical inference-time scaling: sampling 8 candidates and selecting the best via the verifier substantially improves success rates over a single candidate.","MoWAM's motion representation can be computed from any robot with known end-effector geometry, suggesting broad applicability beyond the simulated benchmarks tested.","The motion verifier can rank alternative action outcomes without executing them, enabling safer test-time search for robot policies."],"supporting_citations":[{"why":"provides the pretained video foundation model that the Video Transformer is initialized from and fine-tuned on robot data.","marker":"[12]"},{"why":"s the baseline that removes future video generation at inference (Fast-WAM), against which MoWAM shows improved OOD robustness.","marker":"[16]"},{"why":"s a WAM baseline with dense future video generation at inference (otus), used as a primary comparison for performance and latency.","marker":"[9]"},{"why":"s another WAM baseline using pretrained video diffusion for joint video-action modeling, demonstrating the cost of full future generation.","marker":"[10]"},{"why":"provides the LBERO benchmark suites used for in-distribution evaluation.","marker":"[17]"},{"why":"provides the LBERO-Plus benchmark with seven types of distribution shifts used for OOD evaluation.","marker":"[18]"},{"why":"s the flow matching approach used to model the conditional motion distribution in the Action-Motion Transformer.","marker":"[26]"}],"fun_headline_variants":["Robot motion prediction outruns video generation for faster policies","MoWAM predicts end-effector motion, not whole scenes, for robot learning","Explicit motion prediction makes robot policies more efficient and robust","Skip the video: robot policy learns from motion prediction alone","Future motion prediction boosts robot policy learning while slashing cost"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that the changing geometry of a small set of fingertip and palm keypoints is a sufficient abstraction of future dynamics for action prediction, discarding all other visual details of the scene.","fun_headline_variants_meta":{"raw":{"variants":["Robot motion prediction outruns video generation for faster policies","MoWAM predicts end-effector motion, not whole scenes, for robot learning","Explicit motion prediction makes robot policies more efficient and robust","Skip the video: robot policy learns from motion prediction alone","Future motion prediction boosts robot policy learning while slashing cost"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"cost":0.000192248,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000192248,"upstream_inference_prompt_cost":9.5088e-05,"upstream_inference_completions_cost":9.716e-05},"total_tokens":1231,"prompt_tokens":884,"completion_tokens":347,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":260}},"tokens_in":884,"tokens_out":347,"cost_source":"openrouter_usage","duration_ms":4590,"image_pages":[],"temperature":1.0,"reasoning_tokens":260,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-20T04:25:27.179948+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Test MoWAM on a manipulation task involving a deformable object (e.g., folding a cloth or squeezing a sponge) where the end-effector geometry stays nearly constant but the object shape changes; if success rate drops sharply compared to a full-video WAM, the motion abstraction is insufficient for such tasks.","supporting_citations":[{"cited_title":"Wan: Open and Advanced Large-Scale Video Generative Models","cited_arxiv_id":"2503.20314","evidence_quote":"provides the pretained video foundation model that the Video Transformer is initialized from and fine-tuned on robot data."},{"cited_title":"Fast-WAM: Do World Action Models Need Test-time Future Imagination?","cited_arxiv_id":"2603.16666","evidence_quote":"s the baseline that removes future video generation at inference (Fast-WAM), against which MoWAM shows improved OOD robustness."},{"cited_title":"Motus: A unified latent action world model,","cited_arxiv_id":null,"evidence_quote":"s a WAM baseline with dense future video generation at inference (otus), used as a primary comparison for performance and latency."},{"cited_title":"World Action Models are Zero-shot Policies","cited_arxiv_id":"2602.15922","evidence_quote":"s another WAM baseline using pretrained video diffusion for joint video-action modeling, demonstrating the cost of full future generation."},{"cited_title":"Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,","cited_arxiv_id":null,"evidence_quote":"provides the LBERO benchmark suites used for in-distribution evaluation."},{"cited_title":"LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models","cited_arxiv_id":"2510.13626","evidence_quote":"provides the LBERO-Plus benchmark with seven types of distribution shifts used for OOD evaluation."},{"cited_title":"Flow Matching for Generative Modeling","cited_arxiv_id":"2210.02747","evidence_quote":"s the flow matching approach used to model the conditional motion distribution in the Action-Motion Transformer."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}