{"id":"076f4cbc-7304-4ee3-871e-da5d3950ccca","arxiv_id":"2607.17651","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HCPG-Flow ranks flow-policy candidate actions with a contact-aware, object-centric distance-reduction score, raising mean success over SAC-Flow on ManiSkill, MetaWorld, and four physical tasks.","lead":"A robot-manipulation paper adds an analytic, no-learning selector to flow-based reinforcement-learning policies. It scores candidate actions by how much they shrink distance to the goal, switching from approaching the object to moving it after contact; simulated and real-robot tests show higher task success and faster completion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-contact progress score relies on an unvalidated rigid-transfer assumption; without slip/force data, the claimed contact-progress mechanism is unsupported.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: the post-contact progress score D(u, ρ_task) in Eq. (8) is only a faithful proxy for task progress if TCP displacement transfers to the object. The paper states this as a 'local approximation' in Section III-C, and the real-robot section reports no force, slip, or object-pose measurements—so the assumption is not empirically checked. This matters because Algorithm 1 selects and weights actions directly by this score; if the premise fails, the causal link between HCPG's selection rule and the observed success improvements is broken, even if the improvements themselves are real. The empirical results are plausible—the ablation (Table III) shows Q-guided K=4 underperforms HCPG, which is independent support for the selector's value—but that does not validate the stated mechanism. A direct correlational check between the progress score and realized object displacement is the minimal experiment that would settle the concern. This does not change the reader's conditional verdict; it makes the key condition explicit: before the central claim is accepted, the contact-progress proxy must be validated on contact-rich and deformable-contact tasks, not merely assumed.","tokens_in":8401,"tokens_out":10553,"duration_ms":102671,"concrete_test":"Use ground-truth object states in the ManiSkill simulator (or external motion capture on the real Franka platform). For the trained HCPG policy, log every control step with c_t=1. Record the executed translation u_t^*, the score D(u_t^*, ρ_task), and the realized object-goal distance reduction Δ_t = d(x_o, x_g) − d(x_o', x_g). Compute the Spearman correlation between D and Δ across contact-phase steps, pooled or per task. Require a significantly positive correlation (e.g., lower 95% CI > 0) separately for rigid-contact tasks (PickCube, PushCube, PullCube, peg insertion) and for the rugby grasp. If the correlation is absent, the post-contact ranking is not a faithful progress proxy and the reported success gains need an alternative mechanistic explanation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"At the core of the claimed mechanism is Eq. (8): after contact, the selector scores candidate translations u_i by D(u_i, ρ_task) = u_i^T ρ_task / ||ρ_task||. Section III-C justifies this as the first-order reduction of object-goal distance, but explicitly 'under the local approximation that TCP displacement transfers to the object.' That approximation is load-bearing: if the object slips, deforms, or is non-rigid, D measures TCP motion toward the goal, not actual object progress, and the ranking can select actions that do not advance the task. The paper provides no contact-force, slip, or object-pose measurements in the real-robot experiments (Section V), and the largest real-robot gain is on rugby-ball grasping, where deformation is plausible. The paper itself disclaims global optimality (Section III-C) but never checks the local premise. Consequently, even if the empirical success improvements are real, the explanation 'contact-progress guidance' is not established; the gains could instead come from the soft embedding, the approach-phase signal, or within-set standardization. This is a correctness risk for the central claimed mechanism, not merely a missing robustness test.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HCPG-Flow, a rollout-time candidate-selection mechanism for flow-based policies built on the SAC-Flow backbone. The method samples K=4 actions from the flow actor, computes a contact phase (approach vs. post-contact), scores each candidate translation by the first-order reduction of a task-relevant distance (Eqs. 5–8), standardizes these scores within the candidate set, and forms a temperature-weighted action embedding (Eqs. 9–10). The actor and critic objectives are left unchanged. The authors report that HCPG-Flow improves across-task mean success over SAC-Flow on ManiSkill (87.2% to 96.7%) and MetaWorld (94.7% to 97.1%), and on four physical tasks raises pooled success from 91.7% to 98.3% while reducing successful completion steps by 17.4%.","tokens_in":8725,"tokens_out":4372,"duration_ms":41769,"significance":"The contribution is potentially valuable: it is a zero-learned-parameter module that can be dropped into an existing flow-policy training loop to improve data collection, and the first-order distance-reduction derivation in Eqs. (5)–(6) is clean and internally correct. The paper is also honest in distinguishing the local surrogate from global optimality, and it provides ablations showing that the gain is not merely from K>1 or Q-guided ranking. The empirical scope includes ten simulated and four real-robot tasks, which is useful breadth. However, the significance is currently limited by the absence of validation of the post-contact rigid-transfer assumption, by the small number of seeds and lack of significance testing, and by unspecified per-task adapter details. With those addressed, the method would be a solid practical contribution.","major_comments":[{"comment":"The post-contact score D(u_i, ρ_task) is justified 'under the local approximation that TCP displacement transfers to the object.' This assumption is load-bearing for the claimed mechanism. In Section V there are no contact-force, slip, or post-contact object-pose measurements; rugby-ball grasping is exactly a case where deformation/slipping is plausible. If the assumption fails, the selector ranks TCP motion toward the goal rather than object progress, and the gains could be due to the soft embedding or standardization rather than 'contact-progress guidance.' Please validate by measuring object velocity relative to TCP during contact, add a simulation variant with slip/deformation, or ablate the contact gate (e.g., fixed approach score vs. HCPG).","section":"§III-C, Eq. (8)"},{"comment":"The headline 9.5-point ManiSkill gain rests on three seeds with no significance tests and is dominated by PickCube, where SAC-Flow has high variance (57.5±14.2 vs. HCPG 91.7±8.8). The across-task mean is computed over task means, not a paired per-task test. With n=3, the evidence is not statistically strong. Report per-task paired differences, confidence intervals, or individual seed values, and ideally add more seeds. The real-robot results also lack trial-level significance testing (15 trials per task, one training run per method).","section":"§IV-A, Table II"},{"comment":"Several free parameters are fixed without sensitivity analysis: c=0.08, λ_a=0.05, K=4, τ=0.7. Table IV only varies K. Since d_c controls the phase switch and τ controls score sharpness, their values could materially affect performance; without sweeps or a clearly stated selection criterion it is hard to rule out tuning on the evaluation tasks. Add sensitivity analysis for τ, d_c, and λ_a at least on a subset of tasks.","section":"§III-D/E, §IV-B"},{"comment":"The task adapter appears to encode substantial per-task semantics (e.g., 'the lift adapter replaces ρ_task with the positive vertical axis', 'articulated tasks use the task axis or goal displacement'). The paper does not specify the adapter for each of the ten tasks, nor the source of the grasp flag and ρ_task in the real-robot observations. Since the adapter is part of the proposed method, missing specifications make the method difficult to reproduce and leave open the possibility that per-task engineering drives the gains. Provide a complete adapter specification table or release code.","section":"§III-B"}],"minor_comments":[{"comment":"Notation inconsistency: the threshold is written as 'dc' in Eq. (7) but as 'd_c' in §III-D and §IV-B.","section":"Eq. (7)"},{"comment":"The row labeled 'Overall / macro mean' mixes two different quantities: pooled success (91.7, 98.3) and macro-mean completion steps (62.4, 51.6). Use separate rows with explicit labels.","section":"Table V"},{"comment":"The 'Final' metric is not defined. Clarify whether it is the success rate at the final checkpoint or the last evaluation episode.","section":"Table III"},{"comment":"The statement that 'all methods use the same training budget' is not enough for reproducibility; report the baseline hyperparameters (learning rates, batch sizes, update frequencies) or cite a released configuration.","section":"§IV-B"},{"comment":"The caption says 'linearly aligning normalized training budgets' but does not explain the alignment procedure; please define it so the learning-progress comparison is interpretable.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"I see no grounds for a novelty concern; the main risks are experimental verification and statistical rigor. I would encourage the authors to address the contact-transfer validation and add more seeds/trials. The self-citation of the SAC-Flow baseline is acceptable since it is the direct backbone being augmented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely neat: instead of ranking flow-policy proposals with a critic, HCPG scores them by a first-order progress estimate along object-centric approach/task directions, with a contact phase switch. The Taylor-expansion derivation is correct and the method adds no learned parameters or auxiliary losses, so it is a lightweight, sensible way to inject manipulation geometry into rollout-time selection. The ablations also do real work—they show K=4 with HCPG beats both single-sample SAC-Flow and Q-guided K=4 on the three tested tasks, which suggests the gain is not just from sampling more candidates.\n\nThe empirical claims are plausible but not fully established. Three seeds with no significance tests is thin, and the headline ManiSkill gain is dominated by PickCube, where SAC-Flow has a 57.5 ± 14.2 mean—huge variance across seeds. The real-robot results are nice to see, but 15 trials per task with no error bars is suggestive, not conclusive. I also share the stress-test concern: the post-contact score assumes TCP displacement transfers to the object, and the paper gives no slip, force, or object-pose data to check that. The authors do explicitly call it a local approximation and disclaim global optimality, so this is not a hidden contradiction, but it does mean the 'contact-progress guidance' mechanism is not directly validated. The gains could partly come from the soft embedding or within-set standardization.\n\nOther soft spots: hyperparameters (K, tau, d_c, lambda_a) are fixed across tasks but there is no report of how they were chosen; no code is shipped; and the 'objectives unchanged' claim is technically true for the losses but the replay distribution shifts because the executed action is a soft blend, which is a minor semantic quibble. The self-cited SAC-Flow backbone is fine—this is an incremental paper building on that baseline.\n\nOverall, this is a solid submission worth taking seriously. The idea is new to the flow-policy literature, the math is clean, and the empirical pattern is consistent across simulation and real hardware. It deserves a rigorous peer review, but the authors should be pushed for code, more seeds with significance tests, and at least one contact-force or slip measurement on the real robot to back up the mechanism. I would cite this if I work on contact-rich manipulation, and I'd put it on the reading group list.","headline":"A clean, lightweight candidate-selection idea for flow policies with plausible gains, but the mechanism is under-validated and the stats are thin.","tokens_in":9187,"tokens_out":1813,"would_cite":true,"duration_ms":20399,"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":"HCPG-Flow shows that ranking flow-policy actions by geometric contact progress beats critic-based ranking, lifting task success by 9.5 points without changing training.","keywords":["robot manipulation","reinforcement learning","flow policies","action selection","contact-rich manipulation","object-centric control","candidate sampling","HCPG-Flow"],"falsifier":"Run HCPG-Flow on a pushing or grasping task with a slippery or deformable object where the tool point can move without moving the object, and check whether the selected actions' directional scores track the object's true displacement. If the object-goal distance does not decrease in proportion to the score, the method's advantage over Q-guided selection should vanish or reverse.","tokens_in":8322,"feed_emoji":"🤖","tokens_out":6313,"duration_ms":58541,"temperature":0.7,"pith_summary":"The paper asks how a robot should choose one action when a flow policy proposes several. It establishes that ranking proposals with an object-centric geometric score—how much each candidate reduces a task-relevant distance, switching from approaching the object to moving it after contact—is a better rollout-time choice than relying on learned value estimates. The selector is analytic, adds no learned parameters or auxiliary losses, and leaves the actor and critic objectives unchanged. Across ten simulated tasks from two standard manipulation benchmarks it raises the success of the matched flow-policy backbone by up to 9.5 percentage points, and across four physical tasks it reaches 98.3% success while reducing successful completion steps by 17.4%.","feed_headline":"Progress-based action picker lifts robot success by 9.5 points","feed_subtitle":"An analytic selector ranks flow-policy proposals by contact geometry, not critic values, and cuts completion steps 17.4%.","key_machinery":"The carrying object is the first-order directional progress score D(u, ρ) = u^T ρ / max(‖ρ‖, ε), which measures how much a candidate translation u reduces the distance along a target direction ρ. The full selector combines an approach term and a task-progress term weighted by a contact phase variable, subtracts a small translational penalty, standardizes scores within the candidate set, and blends proposals with a temperature-controlled softmax. It replaces critic-based ranking at rollout time and converts a flow policy's multimodal proposals into one executable action.","core_discovery":"The central claim is that candidate selection in flow policies can be decoupled from value estimation. HCPG-Flow uses state geometry to compute a contact-aware progress score: before contact it rewards translations that reduce tool-to-object distance, and after contact it rewards translations aligned with the object-to-goal direction. The score is a first-order distance-reduction surrogate, standardized within the candidate set, and converted into a temperature-controlled action embedding. With the learning objectives untouched, this rollout-time selector alone accounts for the reported gains; the largest absolute gain appears on a pick-and-place-style task where the phase gate must distingu","pith_inferences":["The task adapters are hand-specified for each task family; a natural extension the paper does not test is learning the approach and task vectors from vision or estimated object pose, which could widen the method's coverage.","Since the score is purely geometric, pairing it with tactile or slip sensing could extend the hierarchy to deformable or slippery objects, where the rigid-transfer assumption breaks.","The same progress signal could be injected as a reward-shaping term during training, although the paper deliberately keeps it as a rollout-time selector; whether that would compound the gains is an open question."],"forward_implications":["Any flow- or diffusion-based actor that can produce multiple proposals can use the same geometric progress criterion to pick one, provided the task exposes object-goal geometry in its state.","Because the selector shapes the replay data, training the identical backbone on HCPG-collected transitions yields better final policies than training on single-sample or Q-ranked data.","The method adds no learned scorer, auxiliary loss, or extra backward pass, so the execution-time overhead is limited to K flow integrations and an analytic weighting.","The reported step reductions on real hardware imply that even when success rates are already saturated, progress-guided selection can shorten task completion time."],"fun_headline_variants":["Contact-aware action selector lifts robot success 9.5 pts","HCPG-Flow picks actions by contact progress, cuts steps 17.4%","Analytic selector improves robot manipulation success 9.5 pts","Contact-progress scoring boosts flow-policy success by 9.5"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that after contact the tool point's displacement transfers to the object, so the post-contact score is a faithful measure of task progress; if the object slips, deforms, or contact is non-rigid, the ranking can select actions that do not actually advance the task.","fun_headline_variants_meta":{"raw":{"variants":["Contact-aware action selector lifts robot success 9.5 pts","HCPG-Flow picks actions by contact progress, cuts steps 17.4%","Analytic selector improves robot manipulation success 9.5 pts","Contact-progress scoring boosts flow-policy success by 9.5"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2750,"prompt_tokens":692,"completion_tokens":2058,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":1979}},"tokens_in":436,"tokens_out":2058,"duration_ms":13150,"temperature":1.0,"reasoning_tokens":1979,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:21:55.819784+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HCPG-Flow on a pushing or grasping task with a slippery or deformable object where the tool point can move without moving the object, and check whether the selected actions' directional scores track the object's true displacement. If the object-goal distance does not decrease in proportion to the score, the method's advantage over Q-guided selection should vanish or reverse.","supporting_citations":[],"review_version":1}