{"id":"a776553f-f97b-4cea-b517-9dbddfbf50c7","arxiv_id":"2508.01082","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A student-teacher RL framework, bootstrapped by contact-implicit trajectory optimization demonstrations, achieves zero-shot sim-to-real pivoting using only vision, force, and proprioception.","lead":"Researchers trained a robot to pivot objects on a table using a three-stage pipeline: offline trajectory optimization generates demonstrations, a teacher policy learns with hidden physics information, and a student policy predicts that information from camera and force data. The system transfers to a real robot without fine-tuning, a useful step toward sample-efficient non-prehensile manipulation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot sim-to-real transfer is the least secure part of the claim: the wall-pivoting task relies on sliding contacts the paper says are hard to model, and no quantitative hardware success rates are reported.","rationale":"The paper's strongest claim is the framework-level assertion of sample-efficient learning and zero-shot transfer. While the hardware video is genuine positive evidence (credit for real-robot results), the claim is broader than what is shown. The most load-bearing condition is the fidelity of the quasi-static SE(2)/MuJoCo contact model for the wall-pivoting task, which involves sliding contact. The authors themselves flag this as the source of a larger sim-to-real gap and admit to task failures from slip. This is not an outside-consensus disagreement but an internally acknowledged weak spot. A concrete hardware evaluation with success rates across the domain-randomization range would either confirm the transfer claim or show it holds only for a narrow subset. The sample-efficiency claim is also not fully visible in the provided text, but the transfer evidence is the more directly undermined by the paper's own statements. Therefore the verdict should remain conditional: the manuscript is promising but needs quantitative transfer results and ideally an ablation isolating the contribution of CITO demonstrations.","tokens_in":4185,"tokens_out":6960,"duration_ms":90575,"concrete_test":"Deploy the wall-pivoting policy on hardware for at least 20 trials, sampling object mass, size, and wall/table friction from the same ranges used in simulation-based domain randomization. Define success as reaching the target orientation without object slip failure, and report success rate and per-trial object-angle error versus the same policy in simulation. Also record student-estimator error for mass and friction. If hardware success is more than 30 percentage points below simulation success, or if the estimator error is significantly higher than in simulation, the zero-shot transfer claim for the wall task is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes zero-shot sim-to-real transfer on several pivoting tasks, with robustness to object physical parameters. The wall-pivoting task is the harder, more contact-rich case, and it inherently involves sliding contact between the object, wall, and table. The paper's own Appendix D.4 states these sliding contacts are challenging to model precisely in simulator (e.g., MuJoCo), leading to a larger sim-to-real gap, and the Limitations section reports occasionally observed slight object slip resulting in task failure. The presented transfer evidence is an orientation-over-time plot for three trials and a video, with no success rates, error bars, or held-out object parameters. Thus the key condition for the central claim—that the simulator used for CITO demonstrations and teacher training is faithful enough for zero-shot transfer across the randomized object distribution—is explicitly not met for the wall task. Without quantitative hardware evaluation, the claim of successful zero-shot sim-to-real transfer is supported only anecdotally for the simpler no-wall task and not robustly for the wall task.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for learning closed-loop pivoting manipulation by combining Contact-Implicit Trajectory Optimization (CITO) with deep reinforcement learning in a student-teacher architecture. CITO generates demonstrations in Step 1, a teacher policy is trained with privileged information and these demonstrations in Step 2, and a student estimator is trained in Step 3 to predict privileged information from a history of proprioceptive, visual, and force observations. The authors claim sample-efficient learning relative to standard RL and zero-shot sim-to-real transfer to physical hardware for several pivoting tasks, with robustness to object physical parameters. The paper includes simulation and hardware experiments, qualitative orientation trajectories over three trials, a supplemental video, and an unusually explicit limitations section.","tokens_in":4355,"tokens_out":3236,"duration_ms":41439,"significance":"If the central claims hold, the paper makes a useful contribution to contact-rich non-prehensile manipulation by showing that dynamically feasible demonstrations from CITO can improve RL sample efficiency and that a student estimator can replace privileged information, enabling zero-shot deployment. The work is honest about its limitations, including quasistatic SE(2) assumptions, convex rigid objects, and observed incipient slip. The real-robot trials and video are valuable supporting evidence, and the explicit limitations section is a strength that should be preserved in revision. However, the quantitative support for the strongest claims is currently thin, especially for the wall-pivoting task.","major_comments":[{"comment":"The central claim of zero-shot sim-to-real transfer is not supported by quantitative hardware evaluation. The visible evidence is an orientation-over-time plot for three trials (Fig. 9) and a video; no success rate, number of repeated trials, standard deviation, or held-out object parameters are reported. Since Contribution 3 claims 'robust manipulation performance against various uncertainties' in real-world experiments, the paper needs a tabulated hardware evaluation with a stated success criterion, trial counts, and object parameter ranges, including the wall-pivoting task.","section":"Abstract; Section 6; Appendix D.4"},{"comment":"For the wall-pivoting task, the manuscript itself states that sliding contact between the object, the wall, and the table is 'challenging to model precisely in simulator (e.g., MuJoCo), leading to a larger sim-to-real gap,' and that 'slight object slip ... resulting in task failure' was observed during deployment. These statements directly qualify the zero-shot transfer claim for the contact-rich wall task. The paper should either provide quantitative evidence that transfer succeeds despite this gap (e.g., success rates, slip counts) or explicitly scope the zero-shot claim to the no-wall task.","section":"Appendix D.4; Section 7 (Limitations)"},{"comment":"The paper reports that policy learning becomes significantly more challenging as the range of domain randomization over table and wall friction increases, because high friction can cause sticking and low friction can cause sliding. This makes the demonstrated robustness contingent on the chosen randomization ranges; the manuscript should report the exact ranges used and justify that they cover the real-world variation claimed in the experiments.","section":"Section 7, fifth limitation"}],"minor_comments":[{"comment":"The observation history and the segmentation feature extraction used by the student estimator are described only informally; please define the input dimensions, temporal window, and network architecture with equations or a table.","section":"Section 3 (Method)"},{"comment":"Figure 9 compares only three trials and does not clearly distinguish simulation from hardware trajectories for each task; adding per-task color coding, trial markers, and a legend would improve readability and support the claimed comparison.","section":"Figure 9"},{"comment":"The statement that prior student-teacher manipulation work assumes constant object size [56,49] would benefit from a brief explanation of which assumptions are made in which reference, so readers can assess the claimed novelty precisely.","section":"Section 2 (Related Work)"},{"comment":"Some citations in the limitations section appear to be used for context rather than specific claims (e.g., [74,19] for CITO extensions); please verify that each citation supports the surrounding sentence.","section":"Section 7 (Limitations)"}],"recommendation":"major_revision","confidential_remarks":"The paper is commendably honest about its limitations, but the gap between the abstract's strong zero-shot transfer claim and the reported hardware evidence is substantial. I would encourage the editor to require a quantitative hardware evaluation before publication, as this is a load-bearing part of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a reasonable integration of ideas that have been floating around the sim-to-real and learning-from-demonstration communities. What's new is bringing CITO-generated dynamically feasible demonstrations into a teacher-student RL pipeline for pivoting, and training a student that estimates privileged parameters from force and vision rather than assuming constant object size. That combination is not in the cited prior work, and it makes sense on paper. The paper also does something I appreciate: it states clearly where the approach struggles, including slip failures and a larger sim-to-real gap for the wall-pivoting task because sliding contacts are hard to model in MuJoCo.\n\nThe strengths are real. The pipeline is sensible; using CITO to generate contact-rich demonstrations is a good way to inject structure into RL exploration. The student estimator uses a temporal history of segmentation and force, which is a credible way to infer object mass and friction. The hardware video and orientation trajectories show the policy actually pivots objects on a real robot. The limitations section is unusually candid.\n\nThe soft spots are mostly about evidence, not design. The key claim of zero-shot sim-to-real transfer is supported by a video and three orientation-over-time curves, with no aggregate success rates, no error bars, and no held-out object parameters reported in the visible text. The paper says the wall task has a larger sim-to-real gap, yet the abstract claims successful sim-to-real transfer without hedging. That's the largest disconnect. I'd want to see success counts across randomized objects, and ideally a comparison of the student estimator's errors against the teacher's tolerance. The absence of code and data also makes the quantitative claims hard to check. These are addressable issues, not fundamental flaws.\n\nThe stress-test note focuses on the wall task's sliding contacts; that concern is valid, but the paper doesn't hide it. The central argument holds as a systems integration paper; the weakness is that the headline claim outruns the reported evidence.\n\nWho should read this? People working on non-prehensile manipulation, demo-guided RL, or sim-to-real. It deserves a serious referee, conditional on the authors providing the missing quantitative evaluation in the full version. I'd send it to review rather than desk reject, but the review should push hard on the sim-to-real numbers.\n\nBest,","headline":"A coherent CITO-to-RL pipeline for pivoting with an honest limitations section; the zero-shot sim-to-real claim is plausible but under-quantified.","tokens_in":4897,"tokens_out":2099,"would_cite":false,"duration_ms":24398,"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":"Optimization-generated demonstrations make reinforcement learning for pivoting manipulation sample-efficient and transferable to real robots.","keywords":["contact-implicit trajectory optimization","non-prehensile manipulation","pivoting","demonstration-guided reinforcement learning","student-teacher learning","sim-to-real transfer","force and vision feedback","privileged information"],"falsifier":"On the wall-pivoting task, run the learned policy on hardware while tracking the object's pose with motion capture, and run the identical policy in MuJoCo with friction coefficients measured from the real table and wall. If the hardware orientation trajectory deviates from the simulator's by more than the spread the policy saw under domain randomization, or if the object loses contact as the paper reports occasionally happening, the zero-shot transfer claim is falsified for that task. A second check is to compare the student estimator's predicted mass and friction against direct measurements; if those are accurate but behavior still diverges, the simulator's sliding-contact model is the weak link.","tokens_in":1625,"feed_emoji":"🤖","tokens_out":1731,"duration_ms":76453,"temperature":0.7,"pith_summary":"This paper tries to show that the data appetite of reinforcement learning for contact-rich manipulation can be fed cheaply by a planner rather than by a human or by millions of random trials. It proposes a three-stage pipeline: contact-implicit trajectory optimization (CITO) generates dynamically feasible pivoting demonstrations; a teacher policy trained with RL and privileged object parameters learns from those demonstrations; and a student estimator recovers the privileged parameters from a history of force, vision, and segmentation observations. The claimed payoff is sample-efficient learning and zero-shot sim-to-real transfer on physical hardware, for objects whose mass and size were not seen during training. A sympathetic reader would care because non-prehensile manipulation normally forces a choice between model-based controllers that need privileged information and learning-based controllers that need enormous datasets, and this framework tries to take the better half of both.","feed_headline":"Optimization demos teach robot pivoting with less data","feed_subtitle":"Planner-generated contact trajectories cut the RL data bill and let the policy estimate object mass, size, and friction from sensors.","key_machinery":"The load-bearing components are the CITO planner used to generate demonstrations, the privileged teacher policy, and the student estimator. CITO, contact-implicit trajectory optimization, is a planner that optimizes trajectories over unknown contact forces rather than prescribing contact sequences in advance, and it produces robot, object, and contact-force trajectories under quasistatic SE(2) contact constraints. The teacher policy is trained with access to ground-truth physical parameters, and the student estimator, built from a CNN and a temporal convolutional network, maps a history of force measurements and segmentation features to those same parameters. The mechanism that ties them together is demonstration-guided RL: the CITO trajectories shape exploration so that the teacher learns with far fewer environment interactions, and the student estimator closes the sim-to-real loop by replacing privileged parameters with predictions from sensors available on real hardware.","core_discovery":"The central claim is that optimization-generated demonstrations, which include contact forces rather than just kinematic poses, are a sufficient source of supervision for a privileged teacher policy, and that the privileged information can be predicted at deployment time from a short temporal window of segmentation images and force/torque readings. On two pivoting tasks, with and without an external wall, the resulting student policy outperforms standard RL baselines in training efficiency and outperforms an MPC baseline in robustness to inaccurate privileged information, and it transfers to hardware without fine-tuning. The paper further claims that the student estimator is not restricted to constant object size, because it infers mass, size, and friction coefficients online rather than assuming them known.","pith_inferences":["If the same pipeline were applied to articulated or deformable objects, the CITO step would need replacing with a planner supporting those dynamics; the RL and estimator stages would likely carry over unchanged.","A testable extension is to vary the number of segmentation views: the paper's SE(2) assumption means a single top-down camera suffices, but adding a second view could lift the planar restriction and let the estimator infer object geometry rather than just dimensions.","The occasional incipient-slip failures suggest that adding a slip-prediction head to the student estimator, or feeding it tactile images, could turn near-misses into recoverable actions, a direction the paper names but does not implement.","The paper shows prediction accuracy plots for the student estimator, but not a causal attribution of hardware failures to estimator error versus dynamics mismatch; that attribution would be a natural next experiment."],"forward_implications":["A robot equipped with these learned policies can pivot objects of new mass and size using only its own joint readings, a camera, and a force/torque sensor, with no per-object system identification.","Because demonstrations come from an automated planner rather than human teleoperation, scaling to new tasks only requires a CITO module for that task, not new human data collection.","The framework should extend to other non-prehensile skills such as pushing or sliding, since the method does not assume task-specific priors beyond the ability to generate dynamically feasible demonstrations.","The student estimator's use of temporal history means the policy can react to changes in contact state, which matters when objects slip or when friction varies across surfaces.","The paper's reported larger sim-to-real gap for sliding contact indicates that transfer quality is bounded by the fidelity of the simulator's friction model."],"supporting_citations":[{"why":"Supplies the CITO planner that generates the dynamically feasible demonstrations under quasistatic SE(2) contact constraints.","marker":"[57]"},{"why":"Provides the student-teacher paradigm for zero-shot sim-to-real transfer that the framework adopts.","marker":"[17]"},{"why":"Offers an alternative CITO method with contact smoothing, cited as evidence that the quasistatic assumption is standard and as a possible demonstration generator.","marker":"[5]"},{"why":"Baseline demonstration-guided RL using kinematically feasible RRT demonstrations, which this work extends to dynamically feasible contact-rich demonstrations.","marker":"[40]"},{"why":"Baseline imitation from human videos, contrasted with automated CITO demonstrations for scalability and contact-force content.","marker":"[41]"},{"why":"Model-based mixed-integer convex optimization for pivoting, representing the privileged-information-dependent approach the paper aims to relax.","marker":"[19]"}],"fun_headline_variants":["CITO demos slash data for robot pivoting","Optimization demos cut RL data for pivoting robots","Force and vision demos make pivoting learning efficient","Sample-efficient pivoting via optimization-guided RL","Robot pivoting learned with fewer trials via CITO demos"],"cache_read_input_tokens":7040,"weakest_assumption_plain":"The entire pipeline assumes the simulator's contact model, especially sliding friction between object, wall, and table, matches the real world closely enough that a policy trained in MuJoCo, and the student estimator trained on its sensor outputs, behave the same on real hardware without any fine-tuning.","fun_headline_variants_meta":{"raw":{"variants":["CITO demos slash data for robot pivoting","Optimization demos cut RL data for pivoting robots","Force and vision demos make pivoting learning efficient","Sample-efficient pivoting via optimization-guided RL","Robot pivoting learned with fewer trials via CITO demos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000777,"raw_usage":{"total_tokens":3416,"prompt_tokens":904,"completion_tokens":2512,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":2449}},"tokens_in":520,"tokens_out":2512,"duration_ms":21918,"temperature":1.0,"reasoning_tokens":2449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:50:48.410084+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the wall-pivoting task, run the learned policy on hardware while tracking the object's pose with motion capture, and run the identical policy in MuJoCo with friction coefficients measured from the real table and wall. If the hardware orientation trajectory deviates from the simulator's by more than the spread the policy saw under domain randomization, or if the object loses contact as the paper reports occasionally happening, the zero-shot transfer claim is falsified for that task. A second check is to compare the student estimator's predicted mass and friction against direct measurements; if those are accurate but behavior still diverges, the simulator's sliding-contact model is the weak link.","supporting_citations":[],"review_version":1}