{"id":"02f111fb-22c3-4ead-83ff-aa418b39090d","arxiv_id":"2507.14721","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hierarchical reinforcement learning framework with a CVAE contact-location model lets a parallel gripper grasp otherwise unreachable objects on tall walls by combining pushing, pivoting, and grasping, with 90% real-world success across six unseen objects.","lead":"This paper tests a robotic system that can grasp flat objects pressed against a wall by first pushing and pivoting them into a better pose, then grabbing them. A planner chooses between three actions using reinforcement learning, and a generative model suggests where on the object to touch.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-world experiments avoid the tall 'grasp constraining' wall that defines the task; Section V-B's equivalence claim is unverified, so the central sim-to-real claim lacks direct support.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: the real-world validation uses a short wall while the problem statement emphasizes tall, grasp-constraining walls. The paper's own Section V-B acknowledges the limitation and bridges it with an unverified equivalence statement. This matters because the central claim is not merely that the framework can pivot and grasp on an accessible wall; it is that the framework can combine pivoting, pushing, and grasping on a wall where pivoting alone leaves the object ungraspable. The simulation evidence is strong and should be credited, but the zero-shot sim-to-real claim for the central scenario is not directly tested. The concern is concrete and fixable: run real tall-wall trials, or provide contact-height measurements and task-level evidence that the short-wall interaction is dynamically identical. Secondary concerns mentioned by the reader, such as the scripted skill ordering imposed by Eq. (1) penalties and unreported action rescaling, further support a conditional verdict, but the wall-height gap is the most load-bearing because it determines whether the real-world results support the paper's core contribution. The suggested concrete test would settle whether the concern lands; if the tall-wall real trials succeed at comparable rates, the central claim would be substantially strengthened.","tokens_in":10247,"tokens_out":2975,"duration_ms":42031,"concrete_test":"Run at least 10 real-world trials on a wall whose height exceeds the upright height of the test object, so that after pivoting the desired grasp surface remains below the wall top and the object is still not directly graspable; use at least the large box and the small bottle from Table III. If the success rate on this tall-wall setup is not comparable to the 90% average reported in Table III, the Section V-B equivalence assertion fails. As a complementary check, instrument the contact height along the wall in both short-wall and tall-wall trials and verify that the interaction region and required skill sequence are identical.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the framework solves occluded grasping on tall 'grasp constraining' walls, zero-shot, in the real world. The simulation results (Table I and Table II) do test that scenario, but the physical experiments do not. Section V-B explicitly states that, because of camera occlusion, the real wall is built from two magnet bases and is therefore short, and it asserts that 'both on a tall wall and a short wall, the actual physical interaction happens on the lower side of the wall. Hence, the performance of the trained RL policy remains unchanged.' This is an assumption, not a measurement. On a short wall, after pivoting, the object may become graspable immediately, so the skill sequence may reduce to pivot-then-grasp and never exercise the constraining-wall regime where additional pushing is required to make the object graspable. If that is what happens, the 90% average success rate in Table III supports a weaker version of the problem statement, not the tall-wall scenario. The load-bearing assertion that the framework transfers to grasp-constraining walls therefore rests entirely on simulation. This does not refute the method, but it means the real-world evidence is not yet tied to the paper's defining setting.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical reinforcement-learning framework for occluded grasping against tall 'grasp constraining' walls. A DQN high-level policy selects among three low-level skills: a SAC-trained pivoting skill, a hand-coded pushing skill, and a hand-coded grasping skill. A CVAE, conditioned on object point clouds and the skill ID, infers contact locations for pivoting and pushing. The policies are trained entirely in simulation on a box-like object and then deployed zero-shot to six real-world objects. The paper reports 100% success on grasp-constraining walls in simulation (Table I), high success rates on unseen simulated objects across wall lengths (Table II), and a 90% average real-world success rate (Table III), concluding that the framework generalizes and transfers to unseen objects without human demonstrations.","tokens_in":10469,"tokens_out":10206,"duration_ms":130531,"significance":"If the claims as stated were fully supported, the paper would be a useful step toward single-arm occluded grasping on constraining walls: it combines primitive skills with a skill-selection layer, uses a skill-conditioned CVAE for contact locations, reports internal simulation ablations, reports real-world per-object success counts, and does not rely on human demonstrations. These are real strengths. However, two load-bearing issues reduce the significance of the results as currently presented: the real-world validation uses a short wall rather than the tall 'grasp constraining' wall that defines the problem, and the high-level policy's skill ordering is almost entirely fixed by hand-written penalties in Eq. (1), so the claimed 'automatic switching' is not actually demonstrated. The paper is not circular in the derivation sense, since no constants are fit to the reported success rates, but the empirical support for the central tall-wall sim-to-real claim is incomplete.","major_comments":[{"comment":"The physical experiments use a short wall built from two magnet bases, and the paper justifies this by asserting that 'both on a tall wall and a short wall, the actual physical interaction happens on the lower side of the wall,' so performance is unchanged. This is an unverified equivalence assumption. In Section III, a grasp-constraining wall is defined as tall, i.e., the object remains ungraspable after pivoting; on a short wall the object may become graspable immediately after pivoting, so the real-world trajectories may never enter the constraining regime in which additional pushing is required. Table III therefore supports generalization to unseen objects on a short wall, not the paper's central tall-wall scenario. The authors should either run real experiments with a tall wall (e.g., repositioning the camera or using fiducial pose estimates to avoid occlusion) or provide logged evidence from the short-wall trials that the constraining-wall regime was actually exercised. If neither is possible, the abstract and conclusion should be revised to state that the tall-wall claim is supported only by simulation.","section":"Section V-B (Table III)"},{"comment":"The reward penalties in Eq. (1) encode a fixed legal action sequence: pivoting is penalized while vpush=0, pushing is penalized once vpush=1, and grasping is penalized unless vpivot*vpush=1. Consequently, in any given state at most one skill (apart from retries of a failed skill) is non-penalized, and the DQN high-level policy is not learning to switch among skills; it executes a hand-authored push-then-pivot-then-grasp script. This conflicts with the abstract and Section I claim that the framework 'automatically switches' between different types of actions. The paper should either retrain or evaluate the high-level policy with the penalties relaxed or removed, or include environments in which the optimal skill order is not fixed, so that learned switching is actually demonstrated. Otherwise the wording should be changed to acknowledge that the ordering is specified by the reward design.","section":"Section IV-A, Eq. (1)"},{"comment":"The CVAE's role is to infer contact locations, yet the final contact point is chosen by two hand-specified ranking rules ('lowest z position' for pivoting, 'largest y position' for pushing) plus an unspecified offset vector, and no ablation or sensitivity analysis is reported for these rules. In addition, the paper states that x_goal is computed from the wall length l but never gives the formula, and the sim-to-real action rescale factor is not reported. Since contact-location selection is one of the four stated difficulties and a central contribution, these omissions prevent the reader from distinguishing the contribution of the CVAE from that of the manual selection rules, and they hamper reproduction. Please report the exact definitions and add ablations, such as random selection from CVAE outputs, varying the offset vector, or removing the ranking rules.","section":"Section IV-C and Section V"}],"minor_comments":[{"comment":"There are several grammar errors in the text, including 'an grasp accessible environment' and 'on an grasp constraining wall'; these should be corrected.","section":"General"},{"comment":"The notation O ∈ R^{n×m} is confusing: the text later refers to a flattened point cloud of length m, and n is used for the number of point clouds and contact locations. Please define n and m explicitly and clarify whether O is a batch of point clouds or a single cloud.","section":"Section IV-C"},{"comment":"The penalty condition in Eq. (1) is ambiguous because of operator precedence; please parenthesize the disjunctions so that the intended conditions are unambiguous.","section":"Section V-A, Eq. (1)"},{"comment":"The dimensions for the gear object, such as '0.062π × 0.026 [m3]' and '0.05452π × 0.026', are unclear; please use explicit radius and height values.","section":"Section V-A, Table II"},{"comment":"The real-world protocol does not state how many attempts were allowed per object or what the failure modes were. Please report the number of trials per object and, where possible, describe the failure cases.","section":"Section V-B, Table III"},{"comment":"The skill-selection frequency plot has no error bars or indication of variance across the 30 episodes; please add this information and clarify whether the frequencies are normalized per episode.","section":"Section V-A, Fig. 5"},{"comment":"The related work discusses [8] as addressing grasp-constraining walls, but no experimental comparison to [8] is included in Section V; please either add such a comparison or soften the implicit claim of superiority.","section":"Section II-A"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the short wall in the real experiments is well founded and is the main reason I cannot recommend acceptance. The concern about Eq. (1) is better characterized as a hand-encoded skill-order script than as circularity: the high-level policy's choices are heavily constrained by the reward design, but no constants are fit to success rates. The paper's use of hand-coded push/grasp primitives is transparent, but the claims of automatic switching and of zero-shot transfer to tall grasp-constraining walls need to be either substantiated with additional experiments or appropriately narrowed. If the authors can add tall-wall real experiments or substantially soften the central claim, a revision is viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, the short version: this paper has a genuinely useful core — a hierarchical policy that learns to pivot, push, and grasp on tall constraining walls, no demonstrations, zero-shot to unseen objects — and its simulation results are striking: 100% where the strongest baselines get 0–13%. The real-world 90% across six objects is solid evidence of transfer, even if the experiments don't quite test what the title promises.\n\nWhat's new: prior pivot-then-grasp work assumes short, grasp-accessible walls. This paper tackles the tall-wall case, where pivoting alone leaves the object ungraspable, and learns a high-level DQN that switches among pivoting, pushing, and grasping. The CVAE contact-location estimator conditioned on skill ID is a reasonable addition. Training without human demos and transferring to bottles and a circular object is a real step past [7] and [10].\n\nSoft spots, in rough order of importance. First, the real-world wall is short. Section V-B justifies this with an assertion that the physical interaction happens on the lower side of the wall, but that is not a measurement. On a short wall the object may become graspable after pivoting alone, so the 90% may not exercise the constraining-wall regime at all. The tall-wall claim rests entirely on simulation. Second, the high-level policy's autonomy is overstated. The Eq. (1) penalties are so strong they essentially hard-code the sequence pivot → push → grasp; the DQN is choosing when to repeat a skill, not really choreographing novel sequences. That is a minor overclaim if the fixed order is the right one, but it should be acknowledged. Third, there is no scripted baseline. Given the hand-coded push/grasp and the near-fixed ordering, a sequence of hard-coded triggers would be a cheap and informative comparison. Also, the action rescale factor and CVAE offset vector are mentioned but not reported, making sim-to-real reproduction harder than it needed to be.\n\nThe math and citations look fine. The simulation results are internally consistent, and the paper does not fit constants to success rates.\n\nRecommendation: worth a serious referee. The central scenario is under-tested in the real world and the autonomy claim needs softening, but the method and the simulation evidence are real. I would engage with it, ask for the scripted baseline and the tall-wall measurement or a clear justification, then likely accept.","headline":"Solid simulation results and a genuine tall-wall twist, but the real-world experiments sidestep the defining scenario and the high-level policy is less autonomous than claimed.","tokens_in":11057,"tokens_out":1442,"would_cite":true,"duration_ms":15238,"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":"This paper shows that a simulation-trained hierarchy of pivoting, pushing, and grasping skills, coordinated by a DQN and a conditional variational autoencoder, solves occluded grasping on tall grasp-constraining walls, reaching 100%…","keywords":["occluded grasping","extrinsic dexterity","hierarchical reinforcement learning","pivot-push-grasp","conditional variational autoencoder","sim-to-real transfer","grasp constraining walls","domain randomization"],"falsifier":"Run the trained policy in the real world on a genuinely tall, grasp-constraining wall (e.g., with a side-mounted camera or a transparent wall so vision is not blocked), repeat 30 trials per object, and compare success rates with the short-wall results. If tall-wall success drops well below the 90% short-wall average, or below the 100% simulation rate, the claim that the framework solves grasp constraining walls is falsified; the same test can be done in simulation by increasing wall height beyond the training range.","tokens_in":9974,"feed_emoji":"🤖","tokens_out":10240,"duration_ms":101088,"temperature":0.7,"pith_summary":"The paper tackles occluded grasping, where an object's primary graspable faces are blocked by the environment, and specifically the harder case of tall 'grasp constraining' walls where reorienting the object by pivoting still leaves it ungraspable. The authors claim that a hierarchical policy—a high-level DQN that selects among a learned pivoting skill, a hand-coded pushing skill, and a hand-coded grasping skill, with a conditional variational autoencoder predicting contact locations—can combine reorientation and repositioning to make such objects graspable, with no human demonstrations. In simulation the method reaches 100% success on constraining walls while all baselines except a point-cloud variant score 0%; in the real world it transfers zero-shot to six out-of-distribution objects with an average success rate of 90%. The significance is that prior extrinsic-dexterity work assumed grasp-accessible walls, so this extends extrinsic dexterity to environments where a single pivot action is not enough.","feed_headline":"90% real-world success for occluded grasping on constraining walls","feed_subtitle":"Simulation-trained pivot, push, and grasp skills transfer to unseen objects with zero human demos.","key_machinery":"The load-bearing mechanism is the three-part hierarchy that separates discrete task selection from continuous control. A DQN high-level policy observes the object pose and wall length and outputs one of three skill IDs (pivoting, pushing, grasping), with a reward that penalizes selecting a completed subtask or grasping too early. The pivoting skill is a SAC policy in continuous end-effector velocity space trained with domain randomization (Gaussian noise on force and pose); pushing and grasping are hand-coded constant-velocity motions. A conditional variational autoencoder (CVAE), conditioned on the object point cloud and the skill ID, reconstructs contact locations from successful rollouts, and the framework picks the lowest reconstructed point for pivoting and the largest-y point for pushing. This lets the policy re-attempt failed pivots and shift the object along the wall until the grasp is accessible.","core_discovery":"The central claim is that occluded grasping on grasp constraining walls is not solved by pivoting alone, but becomes solvable when the robot can autonomously sequence pivoting, pushing, and grasping, and can choose contact locations from object geometry. Concretely, the paper reports that on tall walls in simulation, the proposed hierarchical framework succeeds in 100% of 30 episodes for boxes and gear objects across wall lengths 0.1–0.4 m, while the SAC baseline, the method without the CVAE, the method without skills, and the Ungraspable baseline all achieve 0% (the Point Cloud baseline achieves 13.3%). In physical experiments, the same simulation-trained policies achieve a 90% average success rate over six unseen objects (boxes, bottles, and a circular object), using an RGB-D camera, AprilTag pose estimation, and admittance control, with no retraining or human demonstrations.","pith_inferences":["A testable corollary the authors do not draw: because the real-world wall was short, the 90% figure does not by itself validate the tall-wall scenario; running the same policy on a genuinely tall wall (with a camera placed to avoid occlusion) would directly test whether the assumed height-invariance of the contact physics holds.","The framework's reliance on hand-coded push and grasp skills and on fixed heuristics for picking contact points (lowest z for pivot, largest y for push) suggests that the CVAE may be doing less work than the hierarchy's other components; a learned selection rule could reveal whether this hand-crafting is the actual bottleneck.","The authors' future-work suggestion of bins with constraining walls is the natural stress test: in a bin the object can contact multiple walls, and the wall-length observation would need to be replaced by a richer representation, so it is unclear whether the same DQN state space would suffice."],"forward_implications":["On grasp-accessible walls the framework matches the best baselines at 100% success, so adopting the hierarchy does not sacrifice performance where simpler pivoting already works.","On grasp-constraining walls the 0% scores of the SAC pivoting baseline and the emergent extrinsic dexterity baseline imply that a single reorientation action is insufficient; the combination of push and pivot is necessary.","Longer walls increase how often the high-level policy selects the pushing skill, showing the skill selector uses the wall-length observation to adapt its strategy rather than executing a fixed sequence.","The CVAE-based contact selection outperforms the point-cloud baseline on constraining walls (100% vs 13.3%), indicating that generative contact-location inference is a critical component for coordinated multi-skill manipulation.","Zero-shot sim-to-real transfer over six unseen objects, including curved bottles and a circle, indicates that the domain randomization and force-based observations make the learned pivoting skill robust to novel geometry."],"supporting_citations":[{"why":"Defines the occluded grasping problem and provides the Ungraspable baseline that assumes grasp-accessible walls and scores 0% on constraining walls.","marker":"[7]"},{"why":"Prior single-arm approach for constraining walls with lower success and a manual skill-switch condition that the paper claims to improve upon.","marker":"[8]"},{"why":"Supplies the SAC pivoting baseline and the pivot-skill formulation the low-level pivoting is built on.","marker":"[9]"},{"why":"Action primitives (push, pivot, grasp) for occluded grasping that the paper contrasts against, since it assumes grasp-accessible walls and box-like objects.","marker":"[10]"},{"why":"HACMAN, the point-cloud contact-location selection approach that inspires the Point Cloud baseline.","marker":"[22]"},{"why":"HACMAN++, the extension that together with [22] defines the Point Cloud baseline the paper compares against.","marker":"[23]"},{"why":"Deep Q-Network, the algorithm used to train the high-level skill-selection policy.","marker":"[25]"},{"why":"Soft Actor-Critic, the algorithm used to train the continuous pivoting skill.","marker":"[26]"},{"why":"AprilTag fiducial system used for object pose estimation in the real-world experiments.","marker":"[27]"}],"fun_headline_variants":["Hierarchical RL with wall contact solves occluded grasping","Pivot, push, grasp: hierarchical RL for occluded objects","Sim-to-real occluded grasping with extrinsics on tall walls","Zero-demo hierarchical RL for occluded grasping on walls","90% real, 100% sim: hierarchical RL for occluded grasping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The real-world experiments use a short wall, and the paper assumes that wall height does not change the physical interaction because contact happens on the lower side; if that assumption is wrong, the 90% success rate does not test the tall-wall scenario that the paper's central claim is about.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical RL with wall contact solves occluded grasping","Pivot, push, grasp: hierarchical RL for occluded objects","Sim-to-real occluded grasping with extrinsics on tall walls","Zero-demo hierarchical RL for occluded grasping on walls","90% real, 100% sim: hierarchical RL for occluded grasping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001142,"raw_usage":{"total_tokens":4781,"prompt_tokens":1025,"completion_tokens":3756,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":3668}},"tokens_in":641,"tokens_out":3756,"duration_ms":32831,"temperature":1.0,"reasoning_tokens":3668,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:49:51.059071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained policy in the real world on a genuinely tall, grasp-constraining wall (e.g., with a side-mounted camera or a transparent wall so vision is not blocked), repeat 30 trials per object, and compare success rates with the short-wall results. If tall-wall success drops well below the 90% short-wall average, or below the 100% simulation rate, the claim that the framework solves grasp constraining walls is falsified; the same test can be done in simulation by increasing wall height beyond the training range.","supporting_citations":[{"cited_title":"Learning to grasp the ungraspable with emer- gent extrinsic dexterity,","cited_arxiv_id":null,"evidence_quote":"Defines the occluded grasping problem and provides the Ungraspable baseline that assumes grasp-accessible walls and scores 0% on constraining walls."},{"cited_title":"Multi-stage re- inforcement learning for non-prehensile manipulation,","cited_arxiv_id":null,"evidence_quote":"Prior single-arm approach for constraining walls with lower success and a manual skill-switch condition that the paper claims to improve upon."},{"cited_title":"Learning generalizable pivoting skills,","cited_arxiv_id":null,"evidence_quote":"Supplies the SAC pivoting baseline and the pivot-skill formulation the low-level pivoting is built on."},{"cited_title":"Learning extrinsic dexterity with parameterized manipulation primitives,","cited_arxiv_id":null,"evidence_quote":"Action primitives (push, pivot, grasp) for occluded grasping that the paper contrasts against, since it assumes grasp-accessible walls and box-like objects."}],"review_version":1}