{"id":"dba82941-f32f-48cd-8b6a-57350c96037e","arxiv_id":"2412.09743","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Consistency-focused planners generate better training data for behavior cloning in contact-rich manipulation than standard sampling-based planners like RRT, enabling zero-shot transfer to hardware.","lead":"The paper tests whether robot manipulation policies can learn from motion-planning demonstrations instead of human teleoperation, and finds that consistent, low-entropy demonstrations are far easier to learn from than diverse RRT demonstrations. The result points to a scalable data-generation strategy for contact-rich dexterous manipulation, with successful zero-shot transfer to real robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Entropy attribution is confounded: RRT vs. greedy/PRM differ in trajectory geometry, regrasp structure, and planner design, not only entropy, and no seeds/error bars support the causal claim.","rationale":"The reader's weakest_assumption correctly identifies the causal attribution problem. I agree that the experiments vary planner design rather than entropy independently, and that the entropy estimates are heuristic proxies. My stress-test sharpens this to specific confounds: the RRT and PRM/greedy pipelines differ in trajectory length, progress distribution, regrasp structure, state coverage, and planner-specific engineering (precomputed primitives, fixed grasps), all of which could plausibly affect BC success. The absence of repeated-seed statistics makes it impossible to assess whether the RRT-vs-greedy gap is even reliable. However, the paper's practical pipeline (low-entropy consistent demonstrations plus diffusion BC) is demonstrated to work well, including zero-shot hardware transfer, so the contribution should not be rejected. The right verdict remains CONDITIONAL: the central claim is plausible but requires a direct entropy manipulation and statistical rigor to be fully established.","tokens_in":13207,"tokens_out":1811,"duration_ms":22879,"concrete_test":"In the IiwaBimanual setup, take the greedy planner (Algorithm 2) and inject controlled stochasticity at a tunable rate: with probability p (e.g., 0, 0.05, 0.2, 0.5), replace the goal-directed PLAN CONTACT subgoal with a uniformly sampled object pose, while keeping all other planner parameters, rollout, and data curation identical. Generate 100-demonstration datasets at each p, train the same diffusion policy with multiple seeds per p, and report success rates with confidence intervals. If success rate declines monotonically with p, entropy is a causal factor; if the gap only appears when switching between completely different planners, the paper's contribution 1 is not supported as stated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Contribution 1 asserts a causal relationship: inconsistent, high-entropy demonstrations degrade BC for contact-rich manipulation. The experiments are correlational. In IiwaBimanual, RRT and greedy search differ in many attributes beyond action entropy: RRT trajectories are longer, contain negative-progress segments, have different regrasp timing and frequency, and explore a broader state distribution. For AllegroHand, the PRM planner not only reduces sampling stochasticity but also replaces the planner itself with precomputed primitives and fixed grasps, changing the demonstration structure entirely. Therefore, the reported performance gap cannot be uniquely attributed to action entropy. Additionally, the entropy metrics are proxies: Shannon entropy over discretized object velocity directions, segment progress, and regrasp Bernoulli events. These are not estimates of p(a|O,g), which is what a BC policy actually learns; a planner could have low velocity-direction entropy yet still produce high conditional action variance given the same observation and goal. Finally, Table I reports single success-rate numbers per dataset size with no seeds, error bars, or hypothesis tests, so even the magnitude of the gap (e.g., 44% vs. 99% at 100 demos) has no stated uncertainty. The load-bearing assumption is that entropy, not other correlated planner properties, drives the degradation. This assumption is never directly tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper investigates whether sampling-based motion planners can serve as data generators for behavior cloning (BC) of contact-rich manipulation skills. It proposes a pipeline that combines a contact-rich planner (RRT, greedy search, or a PRM variant) with simulation rollouts, hindsight goal relabeling, and a diffusion-based goal-conditioned policy. The central empirical claim is that RRT-generated demonstrations have high action entropy, measured through proxy statistics, and that this degrades BC performance, whereas more consistent demonstrations from greedy search (IiwaBimanual) or a PRM-based planner (AllegroHand) yield substantially better policies, including zero-shot hardware transfer. The paper reports success rates for multiple dataset sizes, entropy visualizations, and hardware evaluations for both tasks.","tokens_in":13454,"tokens_out":3851,"duration_ms":45468,"significance":"If the central claim held, the paper would provide actionable guidance for scaling BC beyond teleoperation: instead of maximizing demonstration coverage, planners should be designed to produce low-entropy, consistent demonstrations. The paper's strengths are its concrete and reproducible pipeline, the use of two challenging contact-rich tasks, the dataset-size sweep in Table I, and genuine zero-shot hardware transfer for both tasks. The planner algorithms are described in sufficient detail to reimplement. However, the causal attribution to action entropy is not established by the current experiments, and several load-bearing comparisons lack uncertainty quantification. The pipeline itself is a useful empirical contribution even if the entropy hypothesis is later refined.","major_comments":[{"comment":"The headline comparison in Table I reports a single success-rate number per planner and dataset size, with no repeated seeds, error bars, or significance tests. Diffusion policy training is stochastic, and the evaluation uses 100 random initial object poses, so the gap between 44% and 99% at 100 demonstrations, and the claimed RRT plateau near 85%, have no stated uncertainty. Please report results over at least three training seeds with mean and standard deviation, or otherwise quantify the variability of the comparison.","section":"Section III-A2, Table I"},{"comment":"The entropy measures are not estimates of the conditional action distribution p(A|O,g) that the BC policy actually learns. The paper computes Shannon entropy of discretized object velocity directions, per-segment progress, and regrasp Bernoulli events; these are aggregate statistics. A planner could have low velocity-direction entropy yet high conditional action variance given the same observation and goal, or vice versa. The causal claim of contribution 1 therefore is not directly tested. I recommend either estimating conditional entropy (e.g., by binning observations and goals and measuring action variance within each bin) or designing an experiment that varies demonstration entropy while holding the planner fixed, such as by resampling or reweighting trajectories from the same planner.","section":"Section III-A2"},{"comment":"The RRT versus greedy/PRM comparisons vary many attributes simultaneously, not just entropy. RRT trajectories are longer, meander, contain negative-progress segments, and have different regrasp timing and frequency; the AllegroHand PRM pipeline additionally replaces the planner with precomputed primitives and fixed grasps, and the final system uses a hybrid policy with an open-hand reset. Any of these differences could explain the policy performance gap. The paper's own discussion in Section IV-C attributes failures to out-of-distribution scenarios not present in training data, which points toward coverage and planner structure rather than entropy per se. Please isolate the entropy variable or reframe contribution 1 as a correlational observation and present the pipeline itself as the main contribution.","section":"Section III-A2 and Section III-B"},{"comment":"The test distribution is narrower than the training distribution: the text states that the randomization range is 'slightly shrunken' at test time so that the training set has broader coverage. This protocol can make a consistent planner look better if the two planners have different state coverage, and it limits the generality of the reported success rates. Please report performance on the full training-range distribution as well, or justify the shrunken test set as the intended deployment distribution.","section":"Section III-A2 and Section IV-C"}],"minor_comments":[{"comment":"There is a typo: 'kept the ssame as' should be 'kept the same as'.","section":"Section III-B.3"},{"comment":"The text says 'the Bernouli event' and 'B = 2'; please correct the spelling to 'Bernoulli' and clarify that B denotes the number of bins.","section":"Section III-A2"},{"comment":"The description of the PRM planner says 'traverse through all other canonical orientations'; please clarify whether the planner visits all 24 canonical orientations or only a subset, since this affects trajectory length and the interpretation of the entropy comparison.","section":"Section III-B.1"},{"comment":"The diffusion training loss is written as MSE(epsilon_k, epsilon_theta(...)); a brief description of the noise schedule and the number of diffusion steps would help reproducibility, beyond the reference to [6].","section":"Section II-B"},{"comment":"The caption says 'We show 100 trajectories for each task to avoid cluttering,' but the panels appear to show many overlapping curves; please state explicitly which curves correspond to which planner in each panel, since panels (d) and (e) are easy to confuse.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The paper's central contribution is framed as a causal claim about action entropy, but the experiments only vary planner design. If the authors cannot provide an entropy-controlled experiment, the paper should be reframed as a pipeline paper with a correlational analysis. The use of largely self-cited planners is understandable in this subfield, but the authors should make the independence of the entropy measurement from the planner implementation clearer. The hardware results are a strong point and should be preserved. The paper is within the scope of the journal and is, on the whole, a useful empirical study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper answers a question worth asking: if we use model-based planners to generate training data for behavior cloning in contact-rich manipulation, what properties of the planner matter? The headline comparison—RRT vs greedy search on a bimanual task, RRT vs a PRM with fixed primitives on in-hand reorientation—delivers a clear, reproducible-in-spirit result: policies trained on the more consistent demonstrations do substantially better, and the best policies transfer zero-shot to hardware at decent success rates. That is a real empirical contribution, and the hardware demos are not something many papers in this area bother to do.\n\nThe paper is also upfront about several limitations: it discusses the data imbalance for the AllegroHand task, the hybrid policy, the failure modes, and the sim-to-real gaps. That honesty is to its credit.\n\nThe soft spot is exactly where the stress-test note lands. The paper claims that 'inconsistent, high-entropy demonstrations degrade policy performance,' and that entropy is the causal variable. But the planners differ in many ways beyond action entropy: RRT trajectories are longer, have different regrasp structure, different state coverage, and likely different path quality. The entropy metrics are proxies (velocity directions, segment progress, regrasp frequency), not estimates of p(a|O,g). And Table I reports single success rates with no seeds or error bars, so the 44% vs 99% gap at 100 demos has no stated uncertainty. The AllegroHand comparison is further complicated by the hybrid policy and pre-computed primitives. So the causal attribution is plausible but not demonstrated. What is demonstrated is that, for these two tasks, the specific planners the authors designed produce better BC training data than the RRT baseline.\n\nThe citation pattern is fine; the self-cited planners are the natural baselines, and the paper draws on [13] for the entropy-consistency principle without overselling the novelty.\n\nWho should read this: people working on data generation for imitation learning, especially for contact-rich tasks. It will also be useful for anyone thinking about what 'data quality' means for diffusion policies. A serious referee should engage with it—the paper deserves peer review, and with a request for seeds, error bars, and an analysis that separates entropy from other planner properties, it could become a solid contribution. I'd take it.","headline":"A useful empirical study on planner-generated demos for BC, but the headline causal claim about entropy is not actually isolated.","tokens_in":13982,"tokens_out":2153,"would_cite":true,"duration_ms":21902,"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":"The paper claims that behavior cloning for contact-rich manipulation can succeed with model-based planner demonstrations only when those demonstrations are consistent and low-entropy, and that greedy and PRM planners designed for…","keywords":["behavior cloning","contact-rich manipulation","sampling-based planning","action entropy","demonstration consistency","diffusion policy","goal-conditioned imitation learning","dexterous manipulation"],"falsifier":"Train the same behavior-cloning policy on a low-entropy planner's demonstrations with calibrated random action noise added to match RRT's measured entropy, holding state coverage and path length fixed; if success does not drop as entropy increases, the paper's causal attribution is wrong.","tokens_in":13010,"feed_emoji":"🤖","tokens_out":5732,"duration_ms":55177,"temperature":0.7,"pith_summary":"This paper asks whether sampling-based motion planners can replace human teleoperation as the source of training data for contact-rich manipulation policies. It argues that they can, but only when the planner produces consistent, low-entropy demonstrations: standard RRT planners generate highly variable action distributions that degrade behavior cloning, especially with few demonstrations. The paper shows that swapping RRT for a greedy planner on a bimanual rotation task, and for a fixed-primitive PRM planner on an in-hand reorientation task, yields policies with near-perfect simulation success and successful zero-shot transfer to hardware. The central claim is that demonstration consistency matters more than state coverage when learning from planned data.","feed_headline":"Consistent planner demos beat RRT demos for robot learning","feed_subtitle":"Behavior cloning from consistent, low-entropy planner data reaches high success and transfers zero-shot to hardware.","key_machinery":"The load-bearing machinery is the data-generation pipeline built around a quasi-dynamic contact planner with smoothed contact dynamics, where each step solves a convex optimization problem (Problem 1) to move the object toward a desired pose. The paper modifies this planner twice to lower demonstration entropy: a greedy search that repeatedly solves the same optimization without stochastic subgoal sampling, and a sparse probabilistic roadmap whose nodes are 24 canonical cube orientations connected by three fixed primitives (PitchPlus90, YawPlus45, YawMinus45). These plans are rolled out in a physics simulator without contact smoothing and labeled with hindsight goal relabeling, and a goal-conditioned diffusion policy (a 1D U-Net with FiLM conditioning) is trained to imitate the resulting state-action chunks. Action entropy is the organizing quantity throughout: the paper measures it as entropy of discretized velocity directions, spread of per-segment goal progress, and regrasp-timing entropy.","core_discovery":"On the paper's own terms, the central discovery is that the entropy of planner-generated demonstrations, not their state-space coverage, determines whether a behavior-cloned policy can learn a contact-rich skill. The authors characterize action entropy through three proxies—discretized object velocity direction, per-contact-segment progress toward the goal, and regrasp timing—and show that RRT-based contact planning produces consistently higher entropy on all three. Policies trained on lower-entropy data from a greedy planner reach 99–100% success on the bimanual IiwaBimanual task with 100–5000 demonstrations, while RRT-trained policies plateau around 84–88% success. For the 16-DoF AllegroHand in-hand reorientation task, a PRM planner built from 24 canonical cube orientations and three fixed rotation primitives supports 68–82% simulation success and 62.5% hardware success, with the bimanual policy reaching 90% hardware success. The paper concludes that prioritizing demonstration consistency while maintaining solution diversity is the key design principle for using model-based planners as behavior-cloning teachers.","pith_inferences":["A testable extension is to filter RRT demonstrations by an entropy threshold: if success improves with filtering, entropy is the causal variable; if not, other planner differences matter.","The consistency-over-coverage principle likely extends beyond planners to any demonstration source, including teleoperation and reinforcement learning, where per-state action variance should be measured and controlled.","The PRM-plus-adjustment-policy recipe suggests a modular design: hard constraints encoded in a low-entropy planner, with a small learned policy handling the variable final approach to the goal."],"forward_implications":["Replacing human teleoperation with model-based planners is viable for contact-rich manipulation if the planner is designed for demonstration consistency.","Low action entropy in demonstrations, not broad state coverage, is the main driver of behavior-cloning success in the low-data regime.","Task-specific planner structure—greedy search for planar tasks and a PRM with fixed primitives for dexterous in-hand tasks—can make planned data learnable.","A diffusion-based goal-conditioned policy trained on such data can transfer zero-shot to hardware, achieving 90% success on the bimanual task and 62.5% on in-hand reorientation."],"supporting_citations":[{"why":"Supplies the data-quality principle the paper builds on: good imitation data offers sufficient coverage while maintaining low entropy.","marker":"[13]"},{"why":"Provides the base Contact RRT planner and smoothed quasi-dynamic contact model that the paper analyzes and modifies.","marker":"[21]"},{"why":"Provides the contact trust region PRM planner with canonical orientations and fixed primitives used for the AllegroHand task.","marker":"[24]"},{"why":"Supplies the goal-conditioned diffusion policy architecture, including action chunking, 1D U-Net, and FiLM conditioning.","marker":"[3]"},{"why":"Underpins the quasi-dynamic time-stepping and convex contact formulation used in Problem 1 of the planner.","marker":"[28]"},{"why":"Supplies the hindsight goal relabeling framework used to label reached states as goals in the training dataset.","marker":"[29]"}],"fun_headline_variants":["Low-entropy planner demos beat RRT demos for contact-rich BC","Demo entropy, not coverage, drives contact-rich policy learning","Consistent planner demos enable zero-shot contact-rich manipulation","RRT demos' entropy hurts behavior cloning for dexterous tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the measured performance differences between planners are caused by action-entropy differences, not by other correlated differences such as path optimality, state-distribution narrowness, or task-specific planner engineering.","fun_headline_variants_meta":{"raw":{"variants":["Low-entropy planner demos beat RRT demos for contact-rich BC","Demo entropy, not coverage, drives contact-rich policy learning","Consistent planner demos enable zero-shot contact-rich manipulation","RRT demos' entropy hurts behavior cloning for dexterous tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1172,"prompt_tokens":901,"completion_tokens":271,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":517,"tokens_out":271,"duration_ms":3214,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:46:18.627885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same behavior-cloning policy on a low-entropy planner's demonstrations with calibrated random action noise added to match RRT's measured entropy, holding state coverage and path length fixed; if success does not drop as entropy increases, the paper's causal attribution is wrong.","supporting_citations":[{"cited_title":"Dexterous contact-rich manipulation via the contact trust region,","cited_arxiv_id":null,"evidence_quote":"Provides the contact trust region PRM planner with canonical orientations and fixed primitives used for the AllegroHand task."},{"cited_title":"A convex quasistatic time-stepping scheme for rigid multibody systems with contact and friction,","cited_arxiv_id":null,"evidence_quote":"Underpins the quasi-dynamic time-stepping and convex contact formulation used in Problem 1 of the planner."},{"cited_title":"Goal-conditioned imitation learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the hindsight goal relabeling framework used to label reached states as goals in the training dataset."}],"review_version":1}