{"id":"3237ac0c-2f54-46e8-ad6a-4a42c9f86806","arxiv_id":"2411.09658","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"MBA is a plug-in module that generates robot actions by first diffusing predicted object pose sequences, then conditioning action diffusion on those poses, improving imitation-learning success rates across simulation and real-world benchmarks.","lead":"This paper presents MBA, a module that first predicts the future pose sequence of an object from camera observations and then uses those predicted poses to guide a diffusion-based robot action generator. The authors report that adding this module to existing policies like DP, DP3, and RISE improves success rates on most of 57 simulated tasks and four real-world tasks, at the cost of about double the inference time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates the 'future motion' component: MBA is only compared to vanilla policies, so gains could come from current-pose conditioning or extra capacity. A static-pose-conditioning control is needed before the central claim holds.","rationale":"The reader's conditional verdict is reasonable: missing code, limited real-world trials, unquantified hyperparameters, and no SPOT comparison all warrant conditions. The most load-bearing issue, however, is more specific than the reader's stated factorization. The factorization p(M,A|O)=p(M|O)p(A|M,O) in Section III is a chain-rule identity and is therefore not itself an assumption; the real assumption is that sampling from p(M|O) and conditioning on it is what drives the reported improvement. That assumption is never isolated experimentally. The paper's own interpretation in §IV-B attributes part of the gain to pose estimation for stationary objects, which directly undercuts the 'future motion' narrative. A minimal control—using the current predicted pose repeated over the prediction horizon—would separate future-motion reasoning from static pose conditioning and from the extra capacity of the added module. If the static-pose control recovers the gains, the central claim would need to be weakened to 'auxiliary object-pose conditioning helps diffusion policies,' which is still useful but not the proposed paradigm. This does not change the overall conditional verdict, but it should be added as an explicit requirement before acceptance.","tokens_in":13916,"tokens_out":6855,"duration_ms":70678,"concrete_test":"Run the same 57-task suite with three variants: (A) MBA as published; (B) same MBA architecture, but the action head is conditioned on a constant sequence formed by repeating the first predicted/estimated object pose for Tm steps (current-pose condition); (C) same MBA with a zero or random motion sequence to control for added capacity. If (B) matches (A) within one standard deviation on the aggregate average and on most task groups, the future-motion component is not the driver; if (A) significantly exceeds (B), especially on tasks with genuinely moving objects, the central claim survives. Report per-task success rates across the three seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The experiments show that adding MBA to DP/DP3/RISE improves average success, but they do not show that the improvement comes from predicting future object motion. All baselines are vanilla policies with no object-pose conditioning and no auxiliary motion prediction, so the reported gains conflate at least: (1) additional parameters/capacity of the motion module, (2) benefits of conditioning on object pose at the current frame, and (3) the specific value of a future pose sequence. The paper partially concedes this in §IV-B, where it says that for stationary objects MBA's prediction 'can also be regarded as a pose estimation in this situation' and credits this pose estimation with enabling localization and gripping. Since many benchmark objects are stationary or nearly so during the decisive phase, a condition formed by repeating the current object pose could plausibly reproduce most of the gain. Without an ablation that feeds the current estimated/observed object pose (repeated over the horizon) instead of the predicted future sequence, the central claim of 'motion before action' is not separated from the weaker claim that auxiliary pose conditioning helps diffusion policies.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MBA (Motion Before Action), a plug-and-play module that first uses a diffusion process to predict a future object pose sequence from observations, then feeds that pose sequence as an additional condition to a second diffusion process that generates robot actions. The module is designed to attach to existing diffusion-based imitation policies such as DP, DP3, and RISE. The authors evaluate MBA in simulation across 57 tasks in Adroit, DexArt, and MetaWorld, and in four real-world tasks, reporting consistent average success-rate improvements over the vanilla baselines, plus a small comparison against the flow-based ATM method. The paper also reports learning-curve improvements and discusses limitations including inference cost, MoCap supervision, fixed object count, and deformable-object pose ambiguity.","tokens_in":1753,"tokens_out":1741,"duration_ms":45560,"significance":"If the central claim is validated, namely that conditioning action generation on predicted future object pose sequences is the cause of the observed improvements, the work would be a useful contribution to imitation learning for manipulation, offering a simple module that could be attached to several existing diffusion-policy architectures. The paper has notable strengths: a broad simulation benchmark (57 tasks, three seeds), real-world tests on four tasks, and a design that is agnostic to the visual backbone. It also explicitly discusses limitations. However, the experimental design does not currently isolate the effect of future-motion conditioning from simpler explanations such as added capacity or current-pose conditioning, and the real-world and ATM comparisons lack statistical grounding. These gaps are load-bearing for the paper's main claim, so significant revision is needed.","major_comments":[{"comment":"The primary claim that predicting future object motion is what improves performance is not isolated by the experiments. All baselines are vanilla policies with no object-pose conditioning and no auxiliary motion-prediction module, so the reported gains conflate at least three factors: (1) the additional parameters and training signal of the motion module, (2) the benefit of conditioning on the object pose at the current frame, and (3) the specific value of a predicted future pose sequence. The paper itself concedes in §IV-B that for stationary objects the predicted sequence 'can also be regarded as a pose estimation in this situation,' and many of the benchmark tasks have largely stationary objects during the decisive contact phase. Without an ablation that feeds the current observed or estimated object pose (repeated over the Tm horizon) into the action head, with matched parameter count, the central claim 'motion before action' is not separated from the weaker claim that auxiliary pose conditioning helps. This is the load-bearing experiment for the paper's main contribution and must be added.","section":"§IV-B, Table I, Table II"},{"comment":"The real-world results are reported as point estimates without confidence intervals, per-trial variance, or significance tests. For example, in Cut Clay the DP3 Pick success drops from 95% to 80% when MBA is added, while DP3 w. MBA improves separation from 10% to 20%; in Open Drawer the 20%-to-55% improvement is based on 40 trials, yet no confidence interval is given. Some table entries are also missing (DP and DP w. MBA for Cut Clay are listed as '-'), and the number of trials for the ATM comparison in Table IV is not stated. Because the real-world evaluation is a central part of the paper's evidence that MBA helps substantially, the absence of uncertainty quantification and trial counts makes the real-world claim weaker than presented.","section":"§V-B through §V-E, Table III"},{"comment":"The comparison to ATM is limited to three tasks, only on the DP baseline, and reports no variance or trial count. The paper concludes that the pose-prediction paradigm yields higher-quality action generation than flow prediction, but the evidence is thin: ATM is a full policy with its own tracking architecture, while MBA is a module inserted into DP; the comparison does not control for training budget, network capacity, or inference details. The stated advantage of MBA is attributed to the 'vision-motion gap' and the inability of flow to capture handle pose or rotational dynamics, but these claims are not quantitatively supported by the three-task table. At minimum, the ATM baseline should be run on the same set of tasks and seeds as the other baselines, with equal step counts and reporting conventions.","section":"§V-F, Table IV"},{"comment":"The factorization p(M,A|O)=p(M|O)p(A|M,O) is presented as the formal basis of the method, but its validity is not examined. Since M is itself generated from O by the first diffusion module, conditioning on M is a form of re-encoding O; the decomposition is only useful if M carries information that is not already captured by the observation features feeding the action head. The paper does not test this assumption directly, for instance by comparing against a condition formed from the current pose estimate, which is the same gap raised in the first major comment. Additionally, the method section states that the object pose is a 9D vector with 6D rotation, but for deformable objects such as bread the 6D pose is not well-defined; the Limitations section acknowledges this, but the method section should flag the scope restriction more prominently because it affects the interpretation of the Put Bread into Pot results.","section":"§III, Eq. (1)–(4)"}],"minor_comments":[{"comment":"In the fourth line of Algorithm 1, the update reads 'Ok' where the context indicates it should be 'Mk'; this is confusing because O is used for observations throughout the paper.","section":"Algorithm 1"},{"comment":"The header 'V eryHard' has a spacing typo and should read 'Very Hard'.","section":"Table I"},{"comment":"The learning curves are described as evidence of higher learning efficiency, but no quantitative metric (e.g., area under the curve or steps-to-threshold) is reported; visual inspection alone is not a rigorous comparison.","section":"Fig. 3"},{"comment":"The inference-time comparison reports DP at 95.98 ms and DP with MBA at 197.50 ms, more than a doubling of latency; this cost is mentioned as a limitation, but the paper should also state whether this affects the real-time control loop in the evaluated tasks.","section":"§V-G"},{"comment":"The table footnotes do not specify the number of trials for the ATM rows in Table IV, nor do they explain why DP and DP w. MBA are absent for Cut Clay in Table III; the text mentions that 2D policies were restricted, but the caption should state this explicitly.","section":"Table III and Table IV"},{"comment":"The 'mathematical consistency' argument that object poses and end-effector poses are in the same space and hence share similar distributions is evocative but informal; a more precise statement of what shared distributional structure is assumed would help readers assess the design choice.","section":"§I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a clearly stated contribution and a substantial amount of experimental work, but the central scientific claim is not yet isolated. The missing ablation (current-pose conditioning with matched capacity) is essential and should be requested before acceptance. I also note that no code or data are released, which limits reproducibility of the simulated and real-world results; given the field's norms, this should be addressed. The comparison to ATM is currently too fragmentary to support the paper's stronger statements about flow-based methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead MBA (arXiv:2411.09658). The core idea is simple and worth knowing: before generating actions with a diffusion head, first generate a future object-pose sequence with a second diffusion module, then condition action prediction on those poses. It is plug-and-play with DP, DP3, RISE. The paper reports consistent gains across 57 simulated tasks and four real-world tasks, including a soft-body task, and compares against ATM, a flow-conditioned baseline. That is a solid amount of evidence for a module that is easy to describe.\n\nWhat is genuinely new: the two-stage cascade with pose-sequence conditioning, as a module rather than a full policy. The physical-consistency argument (object poses and end-effector poses share similar space) is reasonable, and the factorization p(M,A|O)=p(M|O)p(A|M,O) is a sensible inductive bias, not a source of circularity — pose supervision comes from MoCap GT, action supervision from expert actions.\n\nThe soft spots are real but not fatal. The biggest is the missing control. The paper claims 'motion before action,' but the baselines are vanilla policies with no pose conditioning at all. The authors themselves note in §IV-B that for stationary objects, the prediction acts as pose estimation. Since many benchmark objects are stationary during the decisive phase, the gains might come from current-pose conditioning or just from having an auxiliary prediction task, not from the 'future' aspect. A simple ablation that feeds the current observed pose (repeated over the horizon) would settle this. Without it, the central claim is not proven.\n\nOther issues are minor by comparison. Real-world results lack confidence intervals and significance tests, although 20 trials per condition is decent. The ATM comparison is limited to three tasks. No code or data is released, and hyperparameters like Tm and Ta are not systematically studied. Inference time roughly doubles (197 ms vs 96 ms for DP), a real cost that the paper acknowledges.\n\nGiven the breadth of experiments and the simplicity of the module, this deserves a serious referee. The referee should require the current-pose ablation, ideally a code/data release, and at least a discussion of statistical uncertainty in the real world. If those are added, the work would be convincing. My own verdict on the evidence as it stands: the general claim that pose conditioning helps is well supported; the specific claim that predicting future motion is the active ingredient is not yet.","headline":"A clean, useful module that conditions diffusion policies on predicted object pose, with broad but not fully controlled experiments; the missing current-pose ablation keeps the central claim from being proven.","tokens_in":14646,"tokens_out":2801,"would_cite":true,"duration_ms":27954,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A robot policy that first predicts the manipulated object's future pose, then generates actions conditioned on that predicted motion, outperforms policies that map observations straight to actions.","keywords":["imitation learning","robotic manipulation","diffusion policy","object motion prediction","6D pose representation","action diffusion","object-centric manipulation","plug-and-play module"],"falsifier":"In any of the reported tasks, swap the predicted object pose sequence at inference for the ground-truth future pose that was used as supervision during training; if success does not rise above the predicted-pose version, the performance gain is not coming from the pose information itself, contradicting the claimed conditioning mechanism.","tokens_in":13740,"feed_emoji":"🤖","tokens_out":8829,"duration_ms":72846,"temperature":0.7,"pith_summary":"This paper proposes that a robot imitation policy should decide what the manipulated object will do before deciding what the robot itself should do. The authors introduce MBA, a plug-in module that runs two diffusion processes in cascade: one predicts the future 6D pose trajectory of the object from the current observation, and the second generates the robot's action sequence conditioned on that predicted motion. They argue this mirrors human motor reasoning and improves both the success rate and the learning speed of existing diffusion-based policies, without changing the policy's inputs or training data. Across 57 simulated tasks and four real-world manipulation tasks, adding MBA raises average success rates over the DP and DP3 baselines by about 14 and 6 percentage points respectively, and it also stabilizes training.","feed_headline":"Adding motion prediction lifts robot task success by 14%","feed_subtitle":"A diffusion module predicts the object's future pose, then guides action generation in 57 simulated and 4 real tasks.","key_machinery":"The central machinery is a cascade of two denoising diffusion probabilistic models sharing a common pose representation: a motion diffusion head that predicts the future 6D pose sequence of the manipulated object from observation features, and the host policy's action diffusion head that generates robot actions conditioned on both the observation and the predicted motion. The load-bearing modeling choice is the factorization $p(M,A|O)=p(M|O)p(A|M,O)$, which turns object motion into a conditioning variable for action generation rather than a separate planner, and it is what lets MBA be inserted into any existing diffusion-head policy without retraining its perception stack.","core_discovery":"MBA models the joint distribution of object motion $M$ and robot action $A$ given observation $O$ as $p(M,A|O)=p(M|O)p(A|M,O)$: a first denoising diffusion process samples the future object pose sequence $M_{t:t+T_m}$ (each pose a 9D vector of 3D translation plus 6D rotation) conditioned on observation features, and a second diffusion process generates the action sequence $A_{t:t+T_a}$ conditioned on both the observation and the encoded predicted motion. The paper argues that because the 6D object pose lives in the same representation space as the robot end-effector pose, the motion is a learnable, physically consistent intermediate variable, so this two-stage cascade yields more accurate and more stable manipulation policies than direct observation-to-action diffusion. The empirical claim is that adding MBA to existing diffusion-head policies improves average success rates by 14.2 points over DP and 6.2 points over DP3 across 57 simulated tasks, and increases success on four real-world tasks including soft-object and articulated-object manipulation.","pith_inferences":["If the pose-conditioning mechanism is the true driver, MBA should extend naturally to multi-object scenes by predicting several pose sequences and concatenating their features; the paper's fixed-size observation vector currently blocks this, a limitation the authors explicitly acknowledge.","The reported doubling of inference time (~197 ms vs ~96 ms for DP) means the success gains come at a lower control frequency; an untested corollary is whether the gains persist when both methods are given the same wall-clock budget per control step.","The paper's success on the soft-bread task is notable because deformable objects are said to lack trackable 6D poses; this suggests the effective conditioning signal may be approximate localization rather than full rigid pose, and distinguishing these would sharpen the claimed mechanism.","A testable extension is to warm-start the motion head with a self-supervised forward model trained on unlabeled video, removing the MoCap supervision cost while keeping the cascade; the paper lists this as future work, not a demonstrated result."],"forward_implications":["Any diffusion-head policy (DP, DP3, RISE) can adopt MBA as a plug-in and is reported to gain success rate, especially on tasks needing precise contact, narrow action spaces, or 6-DoF rotation control.","MBA policies reach their peak success earlier in training and show lower variance across seeds, implying that fewer demonstrations or fewer training iterations may be needed for the same performance.","Because object pose and end-effector pose share the 9D/10D representation, the motion head can reuse the action head's encoding architecture and be trained with the same diffusion objective, simplifying joint optimization.","Conditioning on predicted 6D pose outperforms conditioning on 2D keypoint flow (ATM) in fine-manipulation benchmarks, suggesting pose-space motion is a more informative condition than visual-space flow for action generation.","At deployment, no motion-capture system is needed; the policy runs end-to-end from observations, so the supervision source for object poses can be replaced (e.g., by pose estimators) without altering the policy."],"supporting_citations":[{"why":"Provides the Diffusion Policy baseline with a 2D diffusion action head that MBA is integrated into and compared against in simulation and real-world experiments.","marker":"[8]"},{"why":"Provides the 3D Diffusion Policy baseline whose point-cloud diffusion action head serves as the main 3D integration target and comparison.","marker":"[53]"},{"why":"Supplies the denoising diffusion probabilistic model machinery used for both object-pose generation and action generation.","marker":"[17]"},{"why":"Defines the 6D rotation representation that composes the 9D object pose vector, aligning object motion with the robot's end-effector pose format.","marker":"[57]"},{"why":"Provides the RISE policy, a real-world 3D diffusion-head baseline that MBA is inserted into for the four real-world tasks.","marker":"[42]"},{"why":"Provides the ATM flow-based method whose 2D point-flow conditioning is the alternative compared against MBA's 6D pose conditioning in Table IV.","marker":"[45]"},{"why":"Supplies the Adroit benchmark with dexterous multi-fingered manipulation tasks for evaluating MBA.","marker":"[30]"},{"why":"Supplies the DexArt benchmark for articulated-object manipulation tasks.","marker":"[2]"},{"why":"Supplies the MetaWorld benchmark with a wide range of gripper manipulation tasks across difficulty levels.","marker":"[50]"}],"fun_headline_variants":["Two-step diffusion: predict object pose, then guide robot hand","Predict object motion first, then act: MBA lifts success 14%","Cascaded diffusion: object motion as action guide for robots","MBA diffuses object pose then actions, improving robot tasks by 14%","Motion before action: cascaded diffusion improves robot policies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the future pose sequence of the manipulated object can be reliably predicted from the current observation, and that conditioning the action on that predicted motion preserves all the information about the correct action — if either fails, the entire two-stage cascade loses its advantage.","fun_headline_variants_meta":{"raw":{"variants":["Two-step diffusion: predict object pose, then guide robot hand","Predict object motion first, then act: MBA lifts success 14%","Cascaded diffusion: object motion as action guide for robots","MBA diffuses object pose then actions, improving robot tasks by 14%","Motion before action: cascaded diffusion improves robot policies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1514,"prompt_tokens":901,"completion_tokens":613,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":538}},"tokens_in":517,"tokens_out":613,"duration_ms":5775,"temperature":1.0,"reasoning_tokens":538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:24:18.715514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In any of the reported tasks, swap the predicted object pose sequence at inference for the ground-truth future pose that was used as supervision during training; if success does not rise above the predicted-pose version, the performance gain is not coming from the pose information itself, contradicting the claimed conditioning mechanism.","supporting_citations":[{"cited_title":"Diffusion policy: Visuomotor policy learning via action diffusion,","cited_arxiv_id":null,"evidence_quote":"Provides the Diffusion Policy baseline with a 2D diffusion action head that MBA is integrated into and compared against in simulation and real-world experiments."},{"cited_title":"3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D Diffusion Policy baseline whose point-cloud diffusion action head serves as the main 3D integration target and comparison."},{"cited_title":"Rise: 3d perception makes real-world robot imitation simple and effective,","cited_arxiv_id":null,"evidence_quote":"Provides the RISE policy, a real-world 3D diffusion-head baseline that MBA is inserted into for the four real-world tasks."},{"cited_title":"Dexart: Benchmarking gen- eralizable dexterous manipulation with articulated objects,","cited_arxiv_id":null,"evidence_quote":"Supplies the DexArt benchmark for articulated-object manipulation tasks."},{"cited_title":"Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the MetaWorld benchmark with a wide range of gripper manipulation tasks across difficulty levels."}],"review_version":1}