{"id":"56998ab7-6230-41c5-b6f5-ac5429174eab","arxiv_id":"2411.12982","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-layer diffusion policy, where a Guider predicts the next contact point and an Actor generates the trajectory toward it under Q-learning guidance, outperforms end-to-end Diffusion Policy on contact-rich manipulation.","lead":"This paper introduces a robot control method that first picks a contact point on an object and then generates the motion trajectory to reach it, using two stacked diffusion models plus a learned value function to guide the motion. It reports about 20% higher success than the standard Diffusion Policy across simulated and real manipulation tasks, and it lets a human steer the robot by specifying contact points.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 20.8% improvement claim is an artifact of best-of-three-variant reporting: no single HDP configuration attains it, and the primary Q-learning variant is negative on Square; the headline must be reported per configuration.","rationale":"The reader's weakest assumption about Eq. 14 is real and is corroborated by the paper's own Square limitation, but it explains why the primary variant fails on one task; it does not address the fact that the headline number is unattainable by any single variant. The arithmetic issue is a necessary condition for the central comparative claim. The method has independent strengths: the No-OC 3D encoder helps, prompt guidance is a new capability with quantitative gains, and the Tilt and real-world results are impressive. Those strengths justify a conditional acceptance pending corrected reporting. I therefore disagree with the reader's choice of weakest assumption, while endorsing the CONDITIONAL verdict.","tokens_in":23169,"tokens_out":7799,"duration_ms":88139,"concrete_test":"Recompute the headline average separately for each fixed HDP variant over the same 20 cells, and re-run the strongest cell comparisons (Square-ph-50, Square-mh-50, Tilt-50) against the original Diffusion Policy configuration from [9] with identical network capacity. If the fixed-variant average for eta=0.001 falls below 20.8%, or becomes negative on Square, the abstract and Tables II-III must be revised to report per-configuration numbers and the claim that HDP 'significantly outperforms' DP in general should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Reconstructing Tables II and III shows that the 'HDP Improvement' row is the per-cell maximum improvement over the reduced Diffusion Policy baseline across three variants: HDP(No-OC), HDP(eta=0), and HDP(eta=0.001). No single configuration defines 'HDP' in the headline. On Square-ph-50, the eta=0.001 variant is below DP (0.74/0.60 vs 0.78/0.62), and the reported 19.4% improvement comes from No-OC, which differs from DP only by the added 3D encoder, not by contact guidance. The paper's own Sec. VI-C admits a 2.1% average decrease for eta=0.001 versus No-OC on Square and attributes it to inadvertent pose-changing touches that Algorithm 1 cannot remove. The baseline is also a reduced-size reimplementation of Diffusion Policy, explicitly weaker than the published one [9]. Therefore the abstract's universal superiority claim is not supported by any fixed configuration and conflates architectural ablation gains with the proposed contact-guided Q-learning mechanism.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hierarchical Diffusion Policy (HDP), a two-level imitation learning method for contact-rich manipulation. A high-level Guider network, modeled as a conditional diffusion process, predicts objective contacts from observations including 3D point clouds; a low-level Actor network, also a conditional diffusion process, generates action sequences conditioned on the predicted contact. A Critic network provides a Q-value used to augment the behavior-cloning loss with a Q-learning term. To construct training contacts, the paper introduces a phased-objective-contact algorithm that deletes contacts associated with object-pose-similar phases, which it identifies as misoperations. Additional contributions include a modified PointNet 3D encoder, snapshot gradient optimization, and prompt guidance for human intervention. The paper reports simulation results on Can, Square, Tilt, and Push-T and real-world results on Move-T and Cloth Unfold, claiming an average improvement of 20.8% over Diffusion Policy.","tokens_in":23255,"tokens_out":5272,"duration_ms":56329,"significance":"If the central claim were supported, the paper would make a useful contribution: decomposing manipulation into contact prediction and contact-conditioned trajectory generation, with a Q-value bonus, is a plausible route to improving imitation learning in contact-rich regimes, and the prompt-guidance mechanism is a genuinely attractive controllability feature. The paper also provides unusually detailed empirical documentation: multiple training-sample counts, three seeds, 50 initial conditions, per-variant ablations, and real-world tests on both rigid and deformable objects. The code and data are promised publicly. However, the headline comparison is not tied to any single configuration, and the paper's own Section VI-C concedes a task where the proposed contact guidance hurts performance. These issues are fixable in revision but the claims as written overstate what the evidence shows.","major_comments":[{"comment":"The headline \"average improvement of 20.8%\" is not attributable to any fixed HDP configuration. The rows labeled \"HDP Improvement\" in Tables II and III are per-cell maxima over three variants, HDP(No-OC), HDP(eta=0), and HDP(eta=0.001). For example, on Square-ph(50), the eta=0.001 variant scores 0.74/0.60, below the Diffusion Policy baseline of 0.78/0.62; the reported 19.4% improvement comes from the No-OC average of 0.74/0.62, not from the proposed contact-guided Q-learning variant. Similarly, the 64.3% improvement on Square-mh(50) comes from No-OC, not from the Q-learning variant. The manuscript also contains an internal inconsistency: the abstract and introduction state 20.8%, while Section VI-C reports an average improvement of 17.0%. The central claim must be reported per configuration and per task, with the averaging rule stated explicitly, and the abstract must not present the best of three variants as a single method's improvement.","section":"Abstract, Tables II and III, Section VI-C"},{"comment":"The comparison target is not the published Diffusion Policy used in the abstract's claim. The note under Table II states: \"We reduce the size of the Diffusion Policy network and use the same backbone network in HDP to accelerate training, resulting in slightly lower performance compared to the results in [9].\" The abstract nevertheless says HDP \"significantly outperforms the existing state-of-the-art imitation learning method Diffusion Policy.\" A comparison against a deliberately weakened reimplementation cannot support a superiority claim over the state of the art unless the published numbers from [9] are also reported and the discrepancy is discussed. At minimum, the claim should be scoped to \"Diffusion Policy with the reduced backbone used here.\"","section":"Table II note and Section VI-B"},{"comment":"The paper's own Section VI-C undermines the load-bearing assumption behind phased objective contacts. It states that for precise tasks such as Square, \"many inadvertent touches that change the object's position which cannot be removed by Algorithm 1, causing the Guider to learn objective contacts that reproduce these mistakes,\" and that contact guidance gives an average decrease of 2.1% on Square when comparing HDP(eta=0.001) to No-OC. Since Eq. (14)'s pose-similarity criterion is what decides which contacts are considered misoperations and are erased from the Guider's targets and the Critic's rewards, this admission means the central preprocessing step can fail on a benchmark included in the headline average. The paper should report results with and without the phased-contact construction on all tasks, discuss when the similarity assumption is invalid, and avoid the universal claim that contact guidance improves performance.","section":"Section VI-C, Algorithm 1, Algorithm 3, Eq. (14)"}],"minor_comments":[{"comment":"The sentence \"Our results show that Diffusion Policy significantly improves state-of-the-art performance across the board\" appears to mean \"HDP\"; as written it contradicts the paper's own comparison. The term \"significantly\" is also used without statistical tests, so it should be replaced with a quantitative statement.","section":"Section VI-C"},{"comment":"The appendix states that for the cloth task \"Algorithm 2 struggles to model the object and accurately compute objective contacts,\" so the authors fall back to immediate contacts. Because the real-world Cloth Unfold result is presented as evidence that HDP handles deformable objects, the main text should state clearly that the proposed phased-contact mechanism was not used for that task and that the reported result is for a degraded version of the method.","section":"Appendix A, Cloth Unfold"},{"comment":"The GitHub URL contains a space, \"https://github.com/dexin-wang/Hierarchical Diffusion Policy\"; it should be a single hyphenated or percent-encoded URL.","section":"Abstract"},{"comment":"The noise augmentation in Eq. (15) has four hyperparameters (mu1, mu2, tau_n1, tau_n2) that are reported only in the appendix tables; a sentence explaining the role of each would improve reproducibility, since the appendix table packs them into a single row.","section":"Section IV-C, Eq. (15)"}],"recommendation":"major_revision","confidential_remarks":"The underlying experimental infrastructure appears sound and the paper is potentially salvageable, but the published abstract and introduction in the current version present a best-of-three-variants result as if it were a single method's improvement. The authors should be asked to recompute all headline numbers per fixed configuration, compare against the actual published Diffusion Policy results or clearly scope the claim, and re-examine whether the universal superiority claim survives the Square and Cloth Unfold caveats. If those issues are addressed, the paper could be a reasonable contribution to imitation learning for contact-rich manipulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for its prompt guidance and its honest ablations, but the headline number is not supported. HDP splits trajectory generation into a Guider that predicts objective contacts and an Actor that generates actions conditioned on those contacts, with a Critic adding a Q-learning objective. The system itself is new, even if the pieces come from existing work. What it does well: extensive evaluation across simulation and real tasks, including rigid and deformable objects; clear algorithms for phased contact extraction; and a candid failure analysis in Sec. VI-C that admits contact guidance hurts on Square. Prompt guidance is the standout contribution—manually specifying contacts steers the robot successfully, with a 145% success-rate gain in Table VI. That is a real capability prior imitation learning methods lack.\n\nThe main problem is reporting. The 20.8% average improvement in the abstract is the mean of per-cell maxima across three HDP variants. No single configuration attains it. On Square, the primary variant HDP(eta=0.001) scores 0.74/0.60 versus the reduced Diffusion Policy baseline's 0.78/0.62, and the paper itself reports a 2.1% average decrease versus its own No-OC variant. The baseline is a reduced-size reimplementation of Diffusion Policy, acknowledged to be weaker than the published model. So the abstract's universal superiority claim is unsupported. The gains from the 3D encoder (No-OC) are conflated with the proposed contact-guided Q-learning mechanism. The paper should report each variant separately and separate architectural gains from contact guidance.\n\nMinor issues: no error bars (though three seeds and max/avg are reported), a malformed code URL with no released artifacts, and several table improvement percentages that require reverse-engineering. The load-bearing assumption that pose-similarity implies misoperation is acknowledged by the authors to fail on Square and cloth, which limits the method's generality.\n\nWho this is for: researchers working on diffusion policies for contact-rich manipulation. The paper deserves peer review; it is coherent, technically detailed, and prompt guidance is a genuine contribution. But the claims need rewriting and the evaluation needs per-configuration numbers with proper statistical reporting. I'd accept for review with the expectation of major revision.","headline":"Coherent contact-guided hierarchical diffusion policy with a genuinely useful prompt-guidance capability, but the 20.8% headline is a best-of-variants artifact and the main Q-learning variant loses on Square; worth reviewing after the claims are recalibrated.","tokens_in":23993,"tokens_out":3408,"would_cite":true,"duration_ms":33962,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Contact guidance lifts robot imitation by 20.8 percent on average across six manipulation tasks.","keywords":["hierarchical diffusion policy","contact guidance","imitation learning","diffusion policy","robot manipulation","Q-learning","prompt guidance","3D point cloud conditioning"],"falsifier":"Take a demonstration set for one task that includes two distinct successful contact strategies whose intermediate object poses are within the thresholds $\\tau_l$ and $\\tau_r$ but whose later outcomes differ. If Algorithm 1 merges these phases and erases one contact, and HDP trained on the merged labels performs worse than HDP trained with both contacts kept, the pose-similarity assumption is falsified.","tokens_in":22755,"feed_emoji":"🤖","tokens_out":5747,"duration_ms":53354,"temperature":0.7,"pith_summary":"Hierarchical Diffusion Policy (HDP) claims that end-to-end diffusion policies underperform on contact-rich manipulation because they never explicitly model where the robot should touch the object. The paper splits the problem into two conditional diffusion processes: a high-level Guider predicts the next objective contact from observations including 3D point clouds, and a low-level Actor generates an action sequence toward that contact, trained by behavior cloning plus a Q-learning term that rewards reaching the contact. Across six simulated and real tasks, the authors report an average 20.8% improvement over Diffusion Policy, with the largest gains on contact-rich and high-precision tasks, plus new capabilities: interpretable subgoals, human prompt guidance, and a faster snapshot-gradient training scheme. If correct, the paper establishes that contact-level subgoals, rather than end poses alone, are a useful decomposition for manipulation imitation learning.","feed_headline":"Contact guidance lifts robot imitation by 20.8%","feed_subtitle":"Two-level diffusion policy: predict where to touch, then generate the trajectory toward that contact.","key_machinery":"The central object is the objective contact $C_t$: the expected contact position between the end-effector and the object at time $t$, or the finger position when already in contact, used as the high-level subgoal. Three networks carry the argument: the Guider, a conditional denoising diffusion process predicting $C_t$; the Actor, a conditional denoising diffusion process predicting the action sequence toward $C_t$; and the Critic, a Q-function whose reward is positive when the robot reaches the objective contact within $T_r$ steps. The phased-objective-contacts algorithm decides which contacts in the demonstrations are real subgoals versus misoperations, and the snapshot-gradient optimization strategy computes the Q-learning gradient at only one denoising timestep per iteration via oneshot denoising (Eq. 11), which the paper reports gives a 4x training speedup and a 24% memory reduction. A modified PointNet 3D encoder, stripped of normalization and input transformations, supplies pose- and size-sensitive object observations. These components together carry the decomposition claim that contact planning and contact-conditioned trajectory generation are learnable as separate diffusion processes.","core_discovery":"HDP establishes that a manipulation policy can be substantially improved by making the predicted contact between robot and object an explicit intermediate variable. The high-level Guider models the multimodal distribution of objective contacts $p(C_t|O_t)$ as a conditional denoising diffusion process conditioned on observations; the low-level Actor models $p(A_t|O_t,C_t)$ and is optimized to minimize diffusion noise prediction error while maximizing the Critic's Q-value for reaching the contact within $T_r$ steps (Eq. 9). A labeling algorithm (Algorithms 1-4) turns raw demonstrations into phased objective contacts and sparse rewards by erasing contacts that do not change object pose, which the authors argue removes misoperations. The authors report that this contact-guided hierarchy outperforms end-to-end Diffusion Policy on all six benchmarks, with an average improvement of 20.8%, and that the resulting trajectories are more interpretable and can be steered in real time by human-specified prompts.","pith_inferences":["Beyond the paper: the contact-as-subgoal formulation suggests a natural path to multi-finger dexterous manipulation, where a single end pose is insufficient but multiple contact points can be predicted jointly; the authors mention this as future work.","Beyond the paper: because the reward and labels are defined by object-pose similarity, the method's success on deformable objects may rely on hand-recorded immediate contacts; a testable extension would be a learned contact-quality model that scores candidate contacts rather than thresholding pose similarity.","Beyond the paper: prompt guidance opens a practical interface for correcting policies online; one could evaluate whether sparse human interventions during deployment improve long-horizon success rates beyond the three-prompt limit tested here.","Beyond the paper: the 20.8% average is across six self-selected tasks; the strongest evidence for the decomposition claim would be a benchmark comparison on contact-rich tasks where the ground-truth misoperations are known, so the labeling assumption can be checked independently."],"forward_implications":["Contact-rich tasks become more learnable: on the Tilt task, contact guidance adds an average 7.9% over the same policy without objective contacts, while the end-to-end baseline saturates at a lower success rate.","Human operators can steer a deployed policy in real time by specifying a prompt contact without retraining; in the Move-T experiment, prompting raised stage-2 success from 29% to 71%.","Decomposing multimodal action distributions into contact planning and contact-conditioned trajectory generation lets a single policy express multiple valid strategies, such as pulling left or pushing right, instead of collapsing onto one mode.","Fewer demonstrations are needed: the gains of HDP over Diffusion Policy are largest in low-sample regimes, and the snapshot-gradient oneshot denoising is more robust when training data is scarce.","The same contact-as-subgoal representation handles rigid and deformable objects and different end-effectors, since contacts generalize across gripper geometry better than end poses do."],"supporting_citations":[{"why":"Supplies the Diffusion Policy baseline and the Actor's convolutional denoising architecture that HDP extends with contact conditioning.","marker":"[9]"},{"why":"Introduces Q-learning for diffusion-based policies, which HDP adapts into the Critic and the snapshot-gradient oneshot denoising.","marker":"[7]"},{"why":"Provides the Tilt task, its reinforcement-learning demonstrations, and the earlier contact-as-subgoal idea that HDP generalizes with a diffusion Guider.","marker":"[30]"},{"why":"Defines the Robomimic Can and Square tasks, the proficient/mixed human demonstration datasets, and the LSTM-GMM baseline.","marker":"[36]"},{"why":"PointNet is the base network the paper modifies into its 3D encoder by removing transformations and normalization.","marker":"[61]"},{"why":"Concurrent hierarchical diffusion policy that predicts end poses instead of contacts; the paper contrasts its pose-based subgoals with contact-based ones.","marker":"[29]"},{"why":"Behavior Transformers baseline used in the simulation comparisons for multimodal action cloning.","marker":"[64]"},{"why":"Supplies the Push-T task adaptation and the energy-based IBC baseline.","marker":"[37]"}],"fun_headline_variants":["Contact prediction guides robot trajectories: 20.8% better","Two-level diffusion policy: predict touch, then move","HDP: Contact guidance boosts imitation by 20.8%","Robots learn where to touch, then how to move"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The labeling algorithm assumes that any contact phase whose ending object pose closely matches the neighboring phase is a misoperation that can be erased; if similar poses can arise from different legitimate contact strategies, the Guider learns to discard valid contacts and the Critic's rewards are built on wrong labels.","fun_headline_variants_meta":{"raw":{"variants":["Contact prediction guides robot trajectories: 20.8% better","Two-level diffusion policy: predict touch, then move","HDP: Contact guidance boosts imitation by 20.8%","Robots learn where to touch, then how to move"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3698,"prompt_tokens":984,"completion_tokens":2714,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2645}},"tokens_in":600,"tokens_out":2714,"duration_ms":18894,"temperature":1.0,"reasoning_tokens":2645,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:01:04.985292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a demonstration set for one task that includes two distinct successful contact strategies whose intermediate object poses are within the thresholds $\\tau_l$ and $\\tau_r$ but whose later outcomes differ. If Algorithm 1 merges these phases and erases one contact, and HDP trained on the merged labels performs worse than HDP trained with both contacts kept, the pose-similarity assumption is falsified.","supporting_citations":[{"cited_title":"What matters in learning from offline human demonstrations for robot manipulation,","cited_arxiv_id":null,"evidence_quote":"Defines the Robomimic Can and Square tasks, the proficient/mixed human demonstration datasets, and the LSTM-GMM baseline."},{"cited_title":"Hierarchical diffusion policy for kinematics-aware multi-task robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"Concurrent hierarchical diffusion policy that predicts end poses instead of contacts; the paper contrasts its pose-based subgoals with contact-based ones."},{"cited_title":"Implicit behavioral cloning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Push-T task adaptation and the energy-based IBC baseline."}],"review_version":1}