{"id":"e588c642-2b64-4eb2-bc29-90214ec80efa","arxiv_id":"2504.18084","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A vision-based dexterous grasping policy trained on RL-generated simulated trajectories mixed with 40 real demonstrations achieves 100% real-world success on all tested objects.","lead":"The authors used a reinforcement-learned policy in a simulated environment to generate 4000 grasping trajectories for a dexterous hand across many superquadric object shapes. When these simulated trajectories were mixed with 40 real teleoperated demonstrations, the resulting vision-based grasping policy succeeded in 15 out of 15 real-world trials, including two objects not used for demonstrations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OOD test objects appear to fall inside the superquadric training distribution, so the claimed generalization to unseen object geometries is not supported by the reported experiments.","rationale":"The reader's weakest_assumption identifies precisely the concern I consider most load-bearing: the OOD objects are convex boxes whose dimensions appear to fall within the sampled superquadric ranges, so the generalization claim rests on an unstated similarity between the synthetic shape distribution and the real test objects rather than on demonstrated coverage of truly novel shape classes. My independent reading of Section IV-B and Table I confirms this: the training shape distribution explicitly includes boxes (epsilon near 2) with sizes in the 2-5 cm by 2-5 cm by 5-10 cm range, and the test objects are box-shaped food packages of comparable size. Without measured parameters, there is no evidence that the test set requires extrapolation in shape space. This directly undermines the paper's headline claim of generalization to 'unseen object shapes' in Section V. The concern is not that the method is wrong, but that the reported experiment does not establish the geometric generalization the paper claims. I also note the secondary issues the reader lists: the real-only OOD success rate is stated as 40% in the text but appears as 10% in Table I, and only five trials per object are run. These reinforce the need for more careful reporting, but they are not as central as the shape-distribution gap. A conditional verdict is appropriate: the method is plausible and the mixed-policy result is promising, but the paper should either demonstrate generalization to objects outside p(phi) or explicitly narrow its claim to within-family interpolation combined with cross-texture and cross-size robustness. Therefore I do not change the reader's conditional verdict.","tokens_in":8741,"tokens_out":6577,"duration_ms":73664,"concrete_test":"Measure the actual superquadric parameters of the three test boxes and compare them to the p(phi) ranges in Section IV-B. Then evaluate the mixed policy on a new set of objects with parameters clearly outside those ranges, such as a flat disk with a3 around 1 cm, a tall cylinder with a3 around 15 cm, and a cone or hexagonal prism, using the same five-trial protocol. If success drops substantially below 100%, the 'unseen object shapes' conclusion must be narrowed to within-family size and texture generalization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that mixing RL-generated simulated trajectories with 40 real demonstrations yields a vision-based policy that generalizes to unseen object geometries. The evidence for this is Table I, whose OOD objects are two boxes: the green jello-size box and the blue spam-size box. However, Section IV-B samples superquadric shape parameters from p(phi) with a1,a2 uniformly in 2-5 cm, a3 in 5-10 cm, and epsilon1,epsilon2 in (epsilon_min, 2]. Since boxes correspond to high epsilon values, and the test boxes appear to fall within these sampled dimensions, they are likely in-distribution with respect to the simulated shape distribution, even if they are out-of-distribution relative to the 40 real demonstrations. The paper never reports the measured superquadric parameters of the test objects, so there is no evidence that the OOD objects exercise geometric extrapolation. The phrase 'unseen object shapes' in Section V is therefore stronger than the evidence supports: the experiment may only demonstrate interpolation within a single shape family under new colors, textures, and sizes. If the mixed policy's 100% success is driven by the 40 real demonstrations plus visual/domain randomization, then the specific contribution of RL-generated geometric diversity is not isolated. This is the load-bearing soft spot: the headline experiment conflates 'unseen relative to real demos' with 'unseen relative to the simulation shape prior.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an RL-driven simulation data-generation pipeline for vision-based dexterous grasping. A parameterized grasp skill consisting of a reference trajectory plus a residual PPO policy is executed in IsaacGym over superquadric object shapes sampled from p(phi), with approach directions sampled from p(z) estimated from real demonstrations, to synthesize training trajectories. These simulated trajectories are mixed with 40 real demonstrations and used to fine-tune Octo-Medium. In real-world tests with one in-distribution object and two out-of-distribution objects, the mixed policy achieves 10/10 OOD successes, while the real-only policy achieves 1/10 and the sim-only policy 0/10; the authors conclude that simulation-augmented data substantially improves generalization to unseen shapes, sizes, and textures.","tokens_in":9031,"tokens_out":3724,"duration_ms":38622,"significance":"The contribution is timely and the two-part control decomposition (reference trajectory plus residual RL policy) is a sensible way to make synthetic dexterous data generation tractable, while the use of privileged information only during data generation is a practical design choice. If the central claim were firmly established, the framework would offer a scalable approach to diversifying dexterous grasping data, and the paper includes a controlled real-hardware comparison across real-only, sim-only, and mixed training. However, the evidence base is thin: each object is evaluated over only five trials, no error bars or statistical tests are provided, the reported OOD success rate for the real-only policy is internally inconsistent, and the paper does not establish that the OOD test objects actually lie outside the simulated shape distribution. The empirical claims are therefore currently stronger than the experiments support.","major_comments":[{"comment":"The text states that the real-only policy \"also generalizes to one of the unseen objects, resulting in a 40% success rate on out-of-distribution cases,\" but Table I reports Total (OOD) = 1/10 (10%). With only one successful OOD trial and five trials per object, the key comparison between real-only and mixed policies rests on a single observation and is not supported by any statistical test. Please correct the inconsistency and provide confidence intervals or additional trials.","section":"Section IV-D, Table I"},{"comment":"The claim that the green and blue boxes are out-of-distribution with respect to the simulated geometry is not established. The object parameterization samples a1 and a2 uniformly between 2–5 cm, a3 between 5–10 cm, and epsilon1,epsilon2 between epsilon_min and 2.0, so box-like superquadrics with high epsilon values fall inside the training distribution if their dimensions lie within these ranges. The paper does not report the measured superquadric parameters of the test objects, so the experiments may only demonstrate interpolation within the sampled shape family under new colors and textures. Please report measured phi values for the test objects, or add test objects from genuinely disjoint shape classes.","section":"Section IV-B, Section IV-D, Fig. 6"},{"comment":"The experimental design does not isolate the contribution of RL-generated geometric diversity. The mixed policy differs from the real-only policy by adding 4000 simulated trajectories that include randomized lighting, camera pose, object pose, and surface textures in addition to shape variation, so the observed OOD improvement could be due to visual and domain randomization rather than the proposed RL-based shape adaptation. An ablation that generates simulated data with object-pose and visual randomization but without the RL residual policy or shape variation (for example, replaying the real demonstrations in simulation) is needed to attribute the gain to the proposed mechanism.","section":"Section IV-D, Table I; Section IV-B"}],"minor_comments":[{"comment":"The word \"simuilator\" is a typo and should read \"simulator.\"","section":"Section IV-D, Training Variants"},{"comment":"The paper does not report the simulation success rate of the RL-generated trajectories, even though Section III-D states that trajectories are added to the dataset only when a success condition is met; reporting this rate would help the reader assess the efficiency and quality of the data-generation pipeline.","section":"Section III-D"},{"comment":"The distribution p(z) is estimated from the same real demonstrations used for the in-distribution test, so the evaluation partially aligns the training prior with the test condition; this alignment should be stated explicitly when interpreting the ID results.","section":"Section IV-B"},{"comment":"The spacing in \"V A models\" is inconsistent with the standard notation \"VA models\" and should be fixed throughout the manuscript.","section":"Abstract and Section I"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read of 2504.18084. The idea is straightforward and worth a look: use a residual RL policy over a parameterized reference trajectory to generate contact-rich grasp trajectories across superquadric object shapes, then mix those with 40 real demos to fine-tune Octo. The modular design is the genuinely useful part—splitting approach from contact and conditioning on skill parameters gives you a lever to align sim data with real behavior. That is a real contribution to the data-generation toolbox, not a revolution.\n\nThe experiments are the problem. The headline is that the mixed policy gets 100% on OOD objects while real-only gets 40% (or 10%—the text and Table I disagree; Table I says 1/10). That inconsistency alone forces a revise. Then there are only five trials per object, no error bars, no ablations on sim data amount or diversity, and no comparison to baseline generators like DexMimicGen or simple replay of the reference trajectory. On the OOD claim, the stress-test holds up: the test objects are boxes, and boxes are high-epsilon superquadrics; the paper samples eps up to 2.0 and a1,a2 in 2–5 cm, a3 in 5–10 cm. If the test boxes are in those ranges, they are in-distribution w.r.t. the sim shape prior, even if unseen in the real demos. The paper never reports the fitted superquadric parameters of the test objects, so “unseen object shapes” is stronger than the evidence supports. If the improvement is mostly due to the 40 real demos plus visual randomization, the contribution of the RL-generated geometric diversity is not isolated.\n\nI want to give credit where earned: the method is described enough to reproduce (PPO in IsaacGym, Octo fine-tuning, sampling ranges given), and the failure of sim-only at 0% is at least an honest sim-to-real gap. The framing is clear. But for a data-generation paper the empirical backbone is thin. No code or data release, no error bars, no shape-parameter measurements of test objects.\n\nWho is it for? People working on dexterous data augmentation or real-to-sim-to-real pipelines will want to know about it, but they should read Table I with a calculator. Would I send it to reviewers? Yes—if an editor believes the contribution is interesting enough, peer review can force the missing measurements and more trials. I'd recommend revision with major asks: statistical power, corrected numbers, OOD objects with fitted superquadrics outside the training distribution, and at least one baseline generator. As is, I'd want to see the author response before relying on the 100% claim.","headline":"A sensible pipeline for mixing RL-generated sim data with a few real demos, but the OOD generalization claim is not yet supported by the evidence—few trials, a 40%/10% inconsistency, and test objects that may fall inside the sim shape prior.","tokens_in":9552,"tokens_out":2623,"would_cite":false,"duration_ms":25571,"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":"Adding RL-generated simulated grasping trajectories to 40 real demonstrations lifts a vision-action policy to 100 percent success on unseen tabletop objects, where real-only training manages 10 percent and sim-only training fails.","keywords":["dexterous manipulation","vision-action models","simulation-to-real transfer","reinforcement learning","imitation learning","data augmentation","superquadrics","residual policy"],"falsifier":"Run the same mixed-data training and real-world evaluation on objects whose superquadric parameters lie outside the sampled ranges, for example a thin disk with $a_3 < 5$ cm, a long rod with $a_3 > 10$ cm, a sphere-like shape with $\\epsilon_1,\\epsilon_2$ near small values, or a non-convex object; if the mixed policy's success rate falls substantially below 100 percent, the claimed generalization is bounded by the simulation prior. Also, training with the same pipeline but withholding the 40 real demonstrations would test how much of the robustness comes from real grounding.","tokens_in":8550,"feed_emoji":"🦾","tokens_out":7074,"duration_ms":63850,"temperature":0.7,"pith_summary":"Vision-action models for dexterous grasping need many demonstrations, but real expert data for multi-finger hands is expensive to collect. This paper tries to break that bottleneck by generating large-scale grasping trajectories in simulation with a reinforcement-learned residual policy, then mixing those trajectories with a small number of real demonstrations to fine-tune a vision-action policy. Across four real tabletop objects, the mixed policy succeeds on all of them, including three unseen shapes, while a policy trained only on the real demonstrations fails on most unseen objects and a simulation-only policy fails entirely. If the claim holds, simulation-augmented data can push dexterous grasp policies toward practical generalization without large-scale real-world data collection.","feed_headline":"RL simulation plus 40 real demos nails unseen-object grasps","feed_subtitle":"Mixed training beats sim-only (0%) and real-only (10%) on out-of-distribution objects in real-world tests.","key_machinery":"The load-bearing object is the parameterized grasp skill $f_{\\mathrm{grasp}}(s_t,z,o)=\\xi_t(z,o)+\\pi_\\theta(s_t)$: a nominal reference trajectory $\\xi_t$ interpolated from a pre-grasp pose to a grasp pose, with the grasp pose estimated from contact-point predictions, plus a residual displacement $\\pi_\\theta$ output by a PPO-trained policy using privileged simulator information. Superquadrics, $\\phi=(a_1,a_2,a_3,\\epsilon_1,\\epsilon_2)$, provide a compact shape space; sampling $z\\sim p(z)$ and $\\phi\\sim p(\\phi)$ yields diverse contact-rich trajectories that are converted into robot-observable state-action pairs and filtered by a lift success condition. This modularity is what lets the generated data stay consistent with human demonstrations while still adapting finger motion to new geometries.","core_discovery":"The paper's central claim is that a dexterous grasping policy can generalize to unseen object geometries when its training set combines a small set of real demonstrations with a much larger set of trajectories generated by an RL-based grasp skill in simulation. The authors report 100 percent success on all four test objects with the mixed policy, versus 10 percent on out-of-distribution objects for the real-only policy and 0 percent for the simulation-only policy. The mechanism is a two-part grasp skill: a parameterized reference trajectory, defined by an approach direction and a grasp pose, plus a residual policy trained with PPO on privileged simulator state such as object pose, contacts, and fingertip forces, which adapts finger motions to each object instance. Object shape is varied through superquadric parameters, and sampling these parameters generates geometrically diverse training data.","pith_inferences":["The generalization claim is demonstrated only on convex box-like objects whose superquadric parameters fall inside the sampled ranges; whether the approach extends to concave, articulated, or otherwise non-superquadric objects is untested.","Because only the final transformer layers are fine-tuned while the vision encoder stays frozen, the measured gains may come mostly from action-space diversity rather than improved visual features; this could be tested by varying the amount of visual augmentation.","A natural scaling experiment would fix the simulation set and reduce the real demonstrations well below 40 to find the minimum real-data budget the mixed policy needs.","The framework could be pointed at other grasp families such as side, pinch, or power grasps by changing the skill-parameter prior, which would test whether approach-direction sampling is the key ingredient behind the reported robustness."],"forward_implications":["Mixing roughly 100 simulated trajectories per real demonstration can push dexterous grasp success from 10 percent to 100 percent on the tested unseen objects.","Simulation-only training transfers poorly in this setting, so real grounding remains necessary for robust vision-based dexterous grasping.","Sampling approach directions from real demonstrations keeps simulated grasps plausible enough to serve as imitation training data.","The reference-plus-residual structure lets the real-to-sim-to-real data-augmentation paradigm work for contact-rich multi-finger manipulation when finger motion is generated by RL rather than reused from demonstrations.","The data-generation pipeline can be tuned through the skill prior $p(z)$ and the superquadric shape prior $p(\\phi)$, making the diversity of generated grasps a controllable knob for downstream policy robustness."],"supporting_citations":[{"why":"Supplies the pretrained 93M-parameter transformer policy that is fine-tuned on the mixed real-and-simulated dataset.","marker":"[27]"},{"why":"Provides the proximal policy optimization algorithm used to train the residual grasp policy.","marker":"[38]"},{"why":"Provides the GPU-based physics simulator where the residual policy is trained and the grasping trajectories are generated.","marker":"[39]"},{"why":"Supplies the contact-point estimation used to compute the grasp pose that anchors the reference trajectory.","marker":"[19]"},{"why":"Provides the reference-trajectory-plus-residual-policy training setup that the paper adapts for data generation.","marker":"[40]"},{"why":"Establishes the real-to-sim-to-real data augmentation paradigm this work extends to dexterous manipulation.","marker":"[10]"},{"why":"Identifies the limitation of reusing finger motions under fixed object shape, which motivates RL-based adaptation in this paper.","marker":"[11]"},{"why":"Motivates the use of privileged simulator information (object pose and contacts) to train the expert policy that generates data.","marker":"[15]"}],"fun_headline_variants":["RL-generated data lifts dexterous grasp generalization to 100%","RL data + real demos: 100% success on unseen objects","Mixed training beats sim-only and real-only for unseen grasps","Simulated RL data + 40 demos yield 100% unseen-object grasp success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the superquadric shapes sampled in simulation are representative of the real objects the policy will encounter; the out-of-distribution test objects are all convex boxes whose dimensions lie within the sampled ranges, so the demonstrated 'unseen object' generalization may not carry over to genuinely novel shape classes.","fun_headline_variants_meta":{"raw":{"variants":["RL-generated data lifts dexterous grasp generalization to 100%","RL data + real demos: 100% success on unseen objects","Mixed training beats sim-only and real-only for unseen grasps","Simulated RL data + 40 demos yield 100% unseen-object grasp success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000584,"raw_usage":{"total_tokens":2710,"prompt_tokens":872,"completion_tokens":1838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1759}},"tokens_in":488,"tokens_out":1838,"duration_ms":12820,"temperature":1.0,"reasoning_tokens":1759,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:24:12.034842+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same mixed-data training and real-world evaluation on objects whose superquadric parameters lie outside the sampled ranges, for example a thin disk with $a_3 < 5$ cm, a long rod with $a_3 > 10$ cm, a sphere-like shape with $\\epsilon_1,\\epsilon_2$ near small values, or a non-convex object; if the mixed policy's success rate falls substantially below 100 percent, the claimed generalization is bounded by the simulation prior. Also, training with the same pipeline but withholding the 40 real demonstrations would test how much of the robustness comes from real grounding.","supporting_citations":[{"cited_title":"Octo: An open-source generalist robot policy,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained 93M-parameter transformer policy that is fine-tuned on the mixed real-and-simulated dataset."},{"cited_title":"Toward automatic robot instruction from perception-mapping human grasps to manipulator grasps,","cited_arxiv_id":null,"evidence_quote":"Supplies the contact-point estimation used to compute the grasp pose that anchors the reference trajectory."},{"cited_title":"Task-grasping from a demonstrated human strategy,","cited_arxiv_id":null,"evidence_quote":"Provides the reference-trajectory-plus-residual-policy training setup that the paper adapts for data generation."},{"cited_title":"Mimicgen: A data generation system for scalable robot learning using human demonstrations,","cited_arxiv_id":null,"evidence_quote":"Establishes the real-to-sim-to-real data augmentation paradigm this work extends to dexterous manipulation."},{"cited_title":"Dexmimicgen: Automated data generation for bimanual dexterous manipulation,","cited_arxiv_id":null,"evidence_quote":"Identifies the limitation of reusing finger motions under fixed object shape, which motivates RL-based adaptation in this paper."},{"cited_title":"Privileged sensing scaffolds reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Motivates the use of privileged simulator information (object pose and contacts) to train the expert policy that generates data."}],"review_version":1}