{"id":"603ff2ea-93b6-46b2-acfa-5ea74de09111","arxiv_id":"2608.01603","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A robot policy that propagates a retrieved contact point through predicted object poses, producing a time-varying affordance trajectory, improves manipulation success over static affordance and pose-only baselines.","lead":"AffordTrajDP improves robot grasping and insertion by turning a fixed \"where to touch\" cue into a moving trajectory that tracks the object as it shifts. It beats three older methods on six simulated tasks and seven real-robot tasks, with the biggest gains on precise insertions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frame ambiguity in Eq. (7): the propagation p_{t+k} = T_{t+k} * c_tgt may be a rigid transform in a fixed frame, not an object-fixed propagation; the paper's own Sec. 5 admits non-accumulation is untested. The central claim is not established.","rationale":"The reader's weakest_assumption correctly focuses on pose accuracy and propagation fidelity. My concern goes further: even if the pose prediction were perfect, the paper's equations do not unambiguously define the frame in which the propagation happens. The central claim—that the propagated affordance is state-consistent—requires that p_{t+k} is computed in an object-fixed frame, but Eq. (7) as written uses an unspecified T_{t+k}. The implementation note about 'target object coordinate frame at time t0' suggests that the poses may be expressed in a fixed initial frame, in which case p_{t+k} = T_{t+k} * c_tgt is just a rigid transform of the initial contact point and does not track the object's evolving pose at all. That would make the method's mechanism inconsistent with its stated contribution. This is not an external disagreement; it is an internal ambiguity that determines whether the central claim is true. The paper's own Sec. 5 admits that the non-accumulation of error is a hypothesis, not a result, which further weakens the claim. I recommend REJECT instead of CONDITIONAL because the frame ambiguity is not a minor fix—it is the core of the method—and because the paper's own text acknowledges the key assumption is untested. A CONDITIONAL verdict would be appropriate if the paper explicitly specified the frame and provided a controlled comparison with injected pose noise; until then, the central claim is not established. I agree with the reader's identification of the pose-accuracy assumption as a weakness, but I think the frame ambiguity is the deeper issue. The empirical results could be explained by the extra conditioning signal rather than by the specific mechanism claimed.","tokens_in":12927,"tokens_out":3450,"duration_ms":29661,"concrete_test":"Re-derive Eq. (7) with the actual frame definitions used in the code/implementation. Specifically: (i) State explicitly what frame T_{t+k} is expressed in (world, target-object-at-t0, target-object-at-t+k, or source-object). (ii) Re-run the ManiSkill3 PegInsertion and PlugCharger experiments with a control variant where the affordance trajectory is replaced by the predicted pose trajectory itself (no contact point propagation) and another variant where the contact point is frozen at its first-frame value. (iii) Plot the true contact location vs. the propagated anchor p_{t+k} under estimated poses at every episode step in PegInsertion; if the mean Euclidean error between p_{t+k} and the true contact point exceeds the task's success tolerance (e.g., 3.77 mm at k=4), then the propagation is not state-consistent and the headline claim is not supported.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that propagating a retrieved contact point through a predicted SE(3) pose trajectory yields state-consistent post-contact guidance. This requires that the predicted poses in Eq. (7) are expressed in the same coordinate frame as the static contact point c_tgt, and that the propagation medium actually tracks the object's evolving pose. Two load-bearing weaknesses: (1) Frame/representation ambiguity. Eq. (4) defines the source-object pose in the target-object frame, but Eq. (7) uses 'T_{t+k}' without specifying its frame. The implementation says that for PickCube 'the target object coordinate frame at time t0 is defined by the pose estimate from FoundationPose at the first observation frame.' If T_{t+k} is a world-frame or t0-frame pose rather than the object's current pose in a fixed frame, then p_{t+k} = T_{t+k} * c_tgt is not a consistent affordance trajectory: it is a rigid transform of the initial contact point, which cannot capture the target object's motion (the very thing the method claims to track). Under the natural reading needed for the claim to hold, p_{t+k} should be (T_obj(t+k))^{-1} * c_tgt or T_{t+k} * c_tgt applied in the object frame, but the paper never states this. (2) The paper's own evidence, Sec. 4.2 and Sec. 5, shows that under estimated history the translation error at k=4 is 3.77 mm, and FoundationPose tracking error rises after contact to a magnitude comparable to the prediction error. More importantly, Sec. 5 explicitly says that whether non-accumulation holds under heavy occlusion or injected pose noise 'remains to be directly validated' and that the evidence is 'a correlational, single-task comparison rather than a controlled ablation.' This is an explicit admission that the central premise—that the propagation medium is accurate enough to preserve contact—is not established. The empirical gains could therefore come from the extra conditioning signal (e.g., the pose trajectory acting as an auxiliary observation) rathe","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AffordTrajDP, a visuomotor policy that augments diffusion-policy imitation learning with a time-varying affordance signal. A static contact point is retrieved from an object memory and transferred to the target via dense visual correspondence; a diffusion model predicts the target object's future SE(3) pose trajectory; the anchor contact point is then propagated through the predicted poses to form an affordance trajectory that conditions a DP3-style action generator. The method is evaluated on six ManiSkill3 tasks and seven real-world tasks on two robot arms, with ablations isolating the contribution of the semantic contact point and the pose-trajectory propagation. The reported results are consistently positive, with the largest gains on precision-critical insertion tasks.","tokens_in":13319,"tokens_out":5597,"duration_ms":63173,"significance":"If the central claim holds, the paper makes a useful conceptual advance: static affordance priors become inconsistent after contact, and jointly preserving 'where to contact' and 'how the contact evolves' is a meaningful, under-specified design axis in affordance-guided manipulation. The paper's strengths include (i) a clean ablation design that separately enables the semantic anchor and the pose-trajectory propagation, (ii) evaluation across two embodiments and seen/unseen object instances, and (iii) an unusually candid discussion in Sec. 5 that labels the non-accumulation of pose-prediction error as a hypothesis rather than an established result. The main weaknesses are the unresolved coordinate-frame definition of the propagation step and the fact that the central 'state-consistent' mechanism is supported only by a single-task correlational error analysis.","major_comments":[{"comment":"The propagation equation p_{t+k} = T_{t+k} c_tgt is underspecified in a load-bearing way. Eq. (4) defines the training target as the source-object pose expressed in the target-object frame, T_src_tgt(t) = (T_tgt_W(t))^{-1} T_src_W(t), which is a relative pose between two objects, not the target object's pose in a fixed frame. The implementation notes then say that for single-interaction-target tasks 'the source and target objects in Eq. (4) both refer to the same object category' and that 'the target object coordinate frame at time t0 is defined by the pose estimate from FoundationPose at the first observation frame.' This does not resolve the frame of T_{t+k} in Eq. (7): is T_{t+k} the object's pose relative to the t0 frame, or its absolute world/camera-frame pose? Is c_tgt expressed in the object frame, the t0 frame, or the camera frame? If T_{t+k} is relative to t0, then p_{t+k} is al","section":"§3.4, Eq. (7), and §4.1 implementation details"},{"comment":"The paper's own Sec. 5 states that the non-accumulation of pose-prediction error is 'a hypothesis for future work rather than an established result' and that the supporting analysis is 'a correlational, single-task comparison rather than a controlled ablation.' This limitation directly bears on the central claim of state-consistent post-contact guidance. Under estimated history, the translation error at k=4 is 3.77 mm, and FoundationPose tracking error rises after contact to a magnitude comparable to the prediction error. Since the propagation mechanism in Eq. (7) inherits exactly these errors, the claim that the affordance trajectory 'tracks the target object's evolving pose' is not yet established for adverse conditions. The authors should either add the noise-injection/occlusion experiments they propose in Sec. 5, or explicitly restrict the central claim to the tested settings. As wri","section":"§4.2, Fig. 6, and §5"},{"comment":"The real-world evaluation reports 30 trials per condition with point counts (e.g., 12/30, 6/30) and no confidence intervals, significance tests, or repeated trials. The paper concludes 'consistently superior success rates' and 'consistently outperforms' from these counts. For example, in Table 3 the difference between AffordDP (25/30) and Ours (29/30) on PickCup is 4 trials; without a binomial confidence interval or a paired protocol this is not statistically distinguishable. The aggregate 'Avg. SR' across 180 or 120 trials mixes tasks and seen/unseen conditions, which further complicates interpretation. I recommend adding confidence intervals or exact tests per condition, or tempering the per-task comparative claims. This is not a fatal flaw, but it is load-bearing for the empirical validation, especially because the real-world results are presented as a headline contribution.","section":"§4.3, Tables 2 and 3"}],"minor_comments":[{"comment":"'Outperforming strong baselines by up to 17.8%' should be '17.8 percentage points' (from 52.2% to 70.0%), not '17.8%'.","section":"Abstract and §1"},{"comment":"The tilde notation c_tgt = T c_src is inconsistent with Eq. (7), where c_tgt is used without a tilde for the homogeneous coordinate. Please unify the notation.","section":"Eq. (2)"},{"comment":"References [3] and [4] appear to be the same paper (3D AffordanceNet, CVPR 2021). Please deduplicate.","section":"References"},{"comment":"The typesetting has missing spaces in several captions (e.g., 'OverviewofAffordTrajDPandcomparisonwithpreviousmethods' in Fig. 1, 'Overview' in Fig. 2). These should be fixed.","section":"Figures and captions"},{"comment":"The discussion phrase 'particularly valuable under limited demonstration data' is not directly evidenced by a data-efficiency plot; consider either adding such an experiment or softening this wording.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a robotics/ML venue and does not appear to contain any circular reasoning or self-citation concerns. The main risk is that the frame ambiguity in Eq. (7) may indicate an actual inconsistency in the method as implemented, not merely an exposition gap; this should be resolved before acceptance. The authors' honest limitation statement in Sec. 5 is commendable and should be preserved, but it also means the central 'state-consistent' mechanism currently rests on weaker evidence than the headline claims suggest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First: this is a real step forward, not a revolution. The synthesis—retrieve a semantic contact point, predict the object's SE(3) trajectory, propagate the point through those poses to get a time-varying affordance—is genuinely new, and the paper shows it buys consistent gains over DP3, AffordDP, and SPOT on ManiSkill3 with the largest margins on insertion tasks. The real-world evaluation spans two arms, seven tasks, seen and unseen instances, which is more breadth than most affordance papers. The ablation is the strongest part: +Contact helps, +Traj alone can actively hurt (the StackCube unseen negative result is a nice control), and only the combination delivers. They also deserve credit for Sec. 5, which frankly admits that the non-accumulation of pose error is a hypothesis, not an established result.\n\nThe soft spots are real but mostly fixable. The biggest is Eq. (7). p_{t+k} = T_{t+k} * c_tgt is underspecified. If c_tgt is expressed in the camera or world frame at t0 and T_{t+k} is the object's pose at t+k, that multiplication is just a rigid transform of a fixed point—it does not track the object's motion. The intended reading must put c_tgt in the object frame (or use a relative pose), and the paper never says so. That ambiguity sits exactly on the load-bearing claim. A referee should demand the frame definitions.\n\nSecond, the paper's own pose analysis undercuts the mechanism story: predicted translation error is 3.77 mm at k=4 under estimated history, and FoundationPose tracking error rises after contact to comparable magnitude. Sec. 5 says the non-accumulation claim is untested under occlusion or injected noise, and the evidence is a correlational single-task comparison. So the task-success numbers support the method; they do not yet support the specific 'state-consistent post-contact guidance' mechanism.\n\nMinor: the real-world results are raw 30-trial counts without statistics; no code or data is released; and the 'necessary and sufficient' wording is broader than three tasks warrant. All addressable.\n\nBottom line: this deserves a serious referee and probably a conditional accept after revision. The core idea is sound, the ablation pattern is consistent, and the authors are unusually honest about what is not yet proven. The revision should fix Eq. (7), add a direct noise-injection or occlusion test, and soften the mechanism claims proportionally.","headline":"Genuinely new synthesis with a clean ablation; the propagation mechanism is plausible, but the frame definition in Eq. (7) and the untested non-accumulation claim keep the 'state-consistent' story from being fully sold.","tokens_in":13872,"tokens_out":4487,"would_cite":true,"duration_ms":46211,"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":"The paper claims that propagating a retrieved contact point through a predicted object-pose trajectory gives state-consistent post-contact guidance, raising six-task simulation success from 52.2% (DP3) and 57.8% (AffordDP) to 70.0% on ManiS","keywords":["affordance-guided manipulation","dynamic affordance trajectory","diffusion policy","object-centric representations","SE(3) pose prediction","imitation learning","robotic manipulation","ManiSkill3"],"falsifier":"Compare AffordTrajDP against AffordDP on an insertion task with injected pose noise of known magnitude added to the tracked history, e.g., 1-4 mm translation and 1-2 degrees rotation, matching the post-contact tracking error the paper measures. If the dynamic-affordance advantage over the static prior disappears or inverts under matched perturbation, then state-consistent propagation is not what is carrying the result.","tokens_in":12855,"feed_emoji":"🤖","tokens_out":6774,"duration_ms":67809,"temperature":0.7,"pith_summary":"Robotic policies that imitate demonstrations often use a static contact point to tell the policy where to touch an object. The paper argues that this prior goes stale as soon as contact nudges the object, so precision tasks like inserting a plug fail. It proposes AffordTrajDP, which predicts how the object's pose will evolve and rigidly carries the contact point along that predicted trajectory, producing contact guidance that tracks the object through the interaction. On six ManiSkill3 tasks the method reaches 70.0% average success versus 52.2% for the point-cloud baseline and 57.8% for the static-affordance baseline, with its largest margins on insertion tasks; seven real-world tasks across two arms reproduce the pattern. The point is that a learned policy should be conditioned not on where contact was, but on where contact will be.","feed_headline":"Moving the contact point with the object lifts robot success to 70%","feed_subtitle":"Dynamic affordance trajectories beat static priors on precision insertions and transfer across two robot arms.","key_machinery":"The central mechanism is the dynamic affordance trajectory τ_{t:t+H} = (p_t, ..., p_{t+H}), with each future contact point obtained as p_{t+k} = T_{t+k} c_target, where c_target is the retrieved anchor contact point and T_{t+k} are predicted SE(3) object poses. The pose trajectory is produced by a DDIM denoiser conditioned on tracked pose history; DDIM's deterministic reverse process keeps the propagated affordance stable across denoising steps. This mechanism turns a once-transferred static contact prior into a moving constraint that stays attached to the object as its pose changes.","core_discovery":"AffordTrajDP's central claim is that the missing ingredient in affordance-guided policies is a conditioning signal that jointly preserves where to contact and how that contact evolves as the object moves. It retrieves a single anchor contact point on a semantically similar source object, transfers it to the target through dense visual correspondence, predicts the target object's future SE(3) poses with a DDIM-based diffusion model conditioned on tracked pose history, and then propagates the anchor through those predicted poses by rigid transformation to form a dynamic affordance trajectory. This time-varying contact prior is fed, alongside the point cloud and proprioception, into a 3D diffus","pith_inferences":["If the paper's observed non-accumulation of pose-prediction error holds under injected noise and heavier occlusion, the same propagation mechanism could keep affordances valid for much longer horizons; the paper itself flags this as future work.","Because propagation is a rigid transform of a single anchor, the method does not require a learned contact model: any future pose estimator with calibrated uncertainty could replace the diffusion trajectory model, making the design modular.","The principle of conditioning on where contact will be rather than where contact was could extend beyond single contact points to oriented affordance frames or grasp poses, and to tasks with actively moving targets.","A sharper test of the paper's necessity claim would be an ablated variant that propagates a semantically ungrounded anchor: the StackCube unseen result already hints such a variant underperforms, but measuring it explicitly would isolate the role of semantic grounding."],"forward_implications":["Static affordance priors become stale once contact moves the object; propagating the anchor through predicted poses prevents the post-contact drift that dominates failures in tasks such as PegInsertion and PlugCharger.","Semantic contact retrieval and pose-trajectory prediction are each necessary but individually insufficient; removing either component drops performance, and pose-only conditioning can even hurt on unseen instances when the anchor is not semantically grounded.","The dynamic affordance signal transfers across embodiments: the same approach improves success over DP3, AffordDP, and SPOT on both a Galaxea A1 arm and a UR7e arm.","Expressing poses in the target object frame makes the guidance layout-invariant, so object placement perturbations and unseen object instances are handled without retraining.","The largest gains appear on the most precision-critical tasks (AdapterInsertion, USB Insertion), suggesting the signal addresses systematic contact-stage failures rather than generic perception noise."],"supporting_citations":[{"why":"Supplies the point-cloud 3D diffusion policy backbone (DP3) that every variant and baseline in this paper builds on.","marker":"[46]"},{"why":"Defines the static affordance prior (contact point plus replayed post-contact trajectory) that AffordTrajDP extends by propagation; also the main comparison baseline.","marker":"[40]"},{"why":"SPOT baseline conditions the policy directly on predicted future SE(3) poses, the alternative signal that lacks contact semantics and is contrasted in the ablation.","marker":"[10]"},{"why":"FoundationPose tracks the object's 6D pose over time, providing the pose history used to train the trajectory model and to propagate the anchor at test time.","marker":"[36]"},{"why":"DDIM provides the deterministic reverse process used to sample predicted pose trajectories without per-step stochastic noise.","marker":"[29]"},{"why":"Any6D reconstructs an object mesh from the initial RGB-D view so that pose trajectories can be extracted from demonstrations.","marker":"[13]"},{"why":"Grounded-SAM segments the object in the first frame, used both for building the affordance memory and for extracting demonstration pose trajectories.","marker":"[25]"},{"why":"CLIP image encoder supplies the appearance embedding used to retrieve the most similar source instance and its anchor contact point.","marker":"[23]"}],"fun_headline_variants":["Dynamic affordances beat static: 70% robot task success","Moving contact point as object moves lifts robot success to 70%","Dynamic affordance trajectories outperform static on robot tasks","Adapting contact via object pose evolution boosts robot success","Time-varying contact priors lift robot manipulation success to 70%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the predicted future object poses are accurate enough that rigidly carrying the contact point through them keeps it on the true contact region; if pose tracking or trajectory prediction drifts, the dynamic guidance drifts along with it, just on a moving target.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic affordances beat static: 70% robot task success","Moving contact point as object moves lifts robot success to 70%","Dynamic affordance trajectories outperform static on robot tasks","Adapting contact via object pose evolution boosts robot success","Time-varying contact priors lift robot manipulation success to 70%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00085,"raw_usage":{"total_tokens":3550,"prompt_tokens":778,"completion_tokens":2772,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2689}},"tokens_in":522,"tokens_out":2772,"duration_ms":22175,"temperature":1.0,"reasoning_tokens":2689,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:11:04.377412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare AffordTrajDP against AffordDP on an insertion task with injected pose noise of known magnitude added to the tracked history, e.g., 1-4 mm translation and 1-2 degrees rotation, matching the post-contact tracking error the paper measures. If the dynamic-affordance advantage over the static prior disappears or inverts under matched perturbation, then state-consistent propagation is not what is carrying the result.","supporting_citations":[{"cited_title":"Spot: Se(3) pose trajectory diffusion for object-centric manipulation, 2025","cited_arxiv_id":null,"evidence_quote":"SPOT baseline conditions the policy directly on predicted future SE(3) poses, the alternative signal that lacks contact semantics and is contrasted in the ablation."},{"cited_title":"Any6d: Model-free 6d pose estimation of novel objects","cited_arxiv_id":null,"evidence_quote":"Any6D reconstructs an object mesh from the initial RGB-D view so that pose trajectories can be extracted from demonstrations."}],"review_version":1}