{"id":"f746aa3f-5186-4a6e-88b1-d787e03c4474","arxiv_id":"2506.15132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Booster Gym provides an end-to-end, open-source RL pipeline with zero-shot sim-to-real transfer, validated on the Booster T1 humanoid for omnidirectional walking, terrain adaptation, and push recovery.","lead":"Booster Gym is an open-source reinforcement learning framework for humanoid robots that trains walking policies in simulation and deploys them to a physical robot, demonstrated on the Booster T1. It packages domain randomization, reward shaping, and a module for parallel ankle mechanisms to make sim-to-real transfer practical.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ankle conversion validation is the key unverified link; without it the central transfer claim lacks quantitative support.","rationale":"The reader identified the same central risk: the kinematic conversion and randomization coverage are unquantified and the evidence is qualitative. I agree with CONDITIONAL because the paper is a useful engineering contribution with a plausible pipeline and a working demo, but the headline claim of seamless zero-shot transfer is stronger than the measured support. My concern is more specifically weighted toward the series-parallel ankle conversion, since it is the one component that is both necessary for the claimed deployment and unvalidated by any experiment. I am not moving the verdict to REJECT because there is no internal inconsistency and the qualitative demo does provide some evidence; I am also not raising it to ACCEPT because the missing quantitative tracking data and the unverified conversion module leave a clear gap. The requested incremental checks, especially the ankle tracking-error experiment and code-level verification of the Jacobian conversion, would be cheap and would directly test the load-bearing assumption.","tokens_in":8798,"tokens_out":1481,"duration_ms":14603,"concrete_test":"Run the released booster_gym deploy code on the real Booster T1 (or on the manufacturer's high-fidelity simulator if available) with the policy frozen, commanding a slow forward walk at 0.5 m/s. Log the virtual-series commands, the converted parallel-ankle commands, and the achieved parallel-ankle joint positions over 200 gait cycles. Compute per-joint mean absolute tracking error and maximum deviation for the ankle pitch and roll axes, and compare against the same metrics when the same policy commands the virtual serial ankle in Isaac Gym under the nominal (unrandomized) parameters. If the real-robot ankle tracking error substantially exceeds the serial-model error, or if the conversion output is not bit-reproducible from the published equations, the zero-shot transfer claim is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is zero-shot sim-to-real transfer on the Booster T1, but the most load-bearing link is the series-parallel ankle conversion described in Section III-D. Because the policy is trained entirely on a virtual serial ankle model, every deployed command passes through the transposed-Jacobian conversion to the physical parallel ankle. If that conversion (or its PD/kinematics implementation) is inaccurate, the whole sim-to-real claim could fail even with perfect domain randomization. The paper reports no quantitative validation of this module: no comparison of commanded versus achieved joint trajectories at the parallel linkage, no conversion-error measurement, no ablation with a directly trained parallel-ankle policy, and no success-rate or tracking-error statistics. The qualitative Fig. 7 shows only that some joint position-velocity trajectories loosely overlap across Isaac Gym, MuJoCo, and the real robot during 0.5 m/s walking; it does not isolate the conversion error. The measured 9-12 ms latency (Section IV-C) and 0-20 ms training randomization do not address kinematic conversion fidelity. This matters because Isaac Gym/PhysX cannot directly simulate the closed kinematic chain, so the conversion module is effectively an untested bridge in the sim-to-real pipeline. An open question is whether the released code actually implements the conversion as described and whether the Jacobian/PD mapping preserves stability and tracking accuracy across the operating range.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Booster Gym proposes an end-to-end reinforcement learning framework for humanoid locomotion, spanning Isaac Gym training, cross-simulator testing in MuJoCo and Webots, and real-world deployment on the Booster T1 robot. The framework uses asymmetric actor-critic PPO with domain randomization, a gait-cycle-conditioned observation, and a series-parallel ankle conversion module in the SDK to reconcile the virtual serial ankle model with the physical parallel ankle mechanism. The paper claims zero-shot sim-to-real transfer and supports the claim with qualitative real-world demonstrations of omnidirectional walking, terrain adaptation, and disturbance recovery, together with a qualitative joint trajectory comparison across Isaac Gym, MuJoCo, and the real robot.","tokens_in":9051,"tokens_out":2755,"duration_ms":30396,"significance":"If the central claim holds, Booster Gym would be a useful community resource: it integrates common RL training components, domain randomization, and a practical solution for parallel ankle mechanisms, with code released as open source. The paper also contributes a concrete latency-measurement procedure and cross-simulator evaluation, and the qualitative hardware demonstrations indicate that the trained policy is at least feasible on the physical robot. However, the evidence currently falls short of supporting the stronger 'seamless transfer' claim: the real-world evaluation is almost entirely qualitative, no repeated-trial statistics or failure rates are reported, the series-parallel conversion module is not quantitatively validated, and the central transfer claim therefore rests on unmeasured assumptions about kinematic conversion fidelity and randomization coverage.","major_comments":[{"comment":"The zero-shot sim-to-real claim depends critically on the series-parallel ankle conversion module, but the paper provides no quantitative validation of this load-bearing component. Specifically, there is no comparison of commanded versus achieved joint trajectories at the parallel linkage, no conversion-error measurement, no ablation with a policy trained directly on a parallel-ankle model, and no success-rate or tracking-error statistics. The qualitative overlap in Fig. 7 cannot isolate the conversion error from other simulation discrepancies, and the acknowledged PhysX limitation with closed kinematic chains (Section II-B) makes this validation necessary.","section":"Section III-D"},{"comment":"The experimental section is almost entirely qualitative, despite the claim in Section IV-C that the dynamics gap is 'systematically quantified.' Only the communication latency and policy inference time are quantified; the actual walking performance, terrain adaptation, push recovery, and cross-simulator fidelity are reported via photographs and single illustrative trajectories, without error bars, repeated trials, failure rates, or baseline comparisons. As a result, the statement in Section V that the framework provides 'zero-shot sim-to-real transfer capabilities' is stronger than the measured evidence.","section":"Sections IV-A, IV-B, IV-C"},{"comment":"Fig. 7 shows joint position-velocity trajectories over a single 0.5 m/s forward-walking condition across Isaac Gym, MuJoCo, and the real robot, but no quantitative similarity metric is computed and no variance information is shown for the 10 Isaac Gym environments. The claim that MuJoCo 'serves as an effective testing platform that closely approximates real-world dynamics' is therefore unsupported; a quantitative comparison over multiple gaits, speeds, and commands is needed.","section":"Fig. 7 and Section IV-C"},{"comment":"The framework acknowledges known simulation weaknesses, particularly the simplified collision estimation in Isaac Gym and the lack of closed-chain support in PhysX, but does not quantify their impact on the final policy. The gait reward substitutes foot height for foot contact forces, and the paper does not test whether this substitution introduces a systematic sim-to-real bias; an ablation or sensitivity study on these modeling choices would materially strengthen the transfer claim.","section":"Section III-B and III-C"}],"minor_comments":[{"comment":"There is a typo in the sentence describing PPO: 'trian' should be 'train'.","section":"Section III-A"},{"comment":"The repository URL appears inconsistently as 'booster_gym' in the abstract and 'booster gym' in the full text; please verify and unify the URL.","section":"Abstract and Section I"},{"comment":"The label 'Parallel-Series' in the figure caption is not defined in the text; the direction of conversion between the parallel and series representations should be clarified.","section":"Fig. 2"},{"comment":"The notation '1 q>qmax + 1 q<qmin' in the joint position limit reward should be defined explicitly as indicator functions.","section":"Table II"},{"comment":"The phrase 'quantitatively quantify' is redundant; consider rewording to 'measure' or 'quantify'.","section":"Section IV-C"},{"comment":"The acknowledgements mention 'legged-gym' and 'rsl-rl' but do not provide explicit citations or repository URLs; these should be included for reproducibility.","section":"Acknowledgements"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about some limitations but the central transfer claim is not supported by the quantitative evidence currently presented. The open-source code availability is central to the paper's contribution, so the editors should verify that the repository is publicly accessible and contains the described conversion module and training configuration. If the code is not available or does not match the described implementation, the verdict may need to be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a solid engineering contribution—an open-source end-to-end RL pipeline for humanoid locomotion, validated on the Booster T1 with real hardware results. If you work on sim-to-real for legged robots, it's worth a look. But the headline claim of 'seamless' zero-shot transfer is stronger than the measurements support.\n\nWhat's actually new: the series-parallel ankle conversion module that lets you train on a virtual serial ankle and deploy on a parallel ankle via a transposed Jacobian and PD mapping, plus the Booster T1 validation. The rest is assembled from known components—PPO, asymmetric actor-critic, domain randomization, gait cycle rewards—and the authors openly acknowledge the open-source repos they drew from. That's fine; the contribution is integration and packaging. The paper does several things well: it measures communication latency (9-12 ms) and randomizes 0-20 ms in training, it tests in two additional simulators (MuJoCo, Webots), and it shows real-world robustness to pushes and terrain variation. Figure 7 shows rough overlap between Isaac Gym, MuJoCo, and real joint trajectories.\n\nThe main weakness is confirmation: no quantitative metrics like tracking error, success rates, or failure counts. The plots are illustrative, not statistical. The stress-test note focuses on the ankle conversion, and that's the right place to worry. The policy is trained entirely on a serial ankle model; every real command is routed through the conversion. The paper gives no direct validation of that module—no comparison of commanded versus achieved parallel joint trajectories, no ablation against a directly trained parallel-ankle policy. So if the Jacobian or PD mapping is off, transfer could degrade or fail even with perfect domain randomization. This is a real gap, but not a fatal one: the hardware demo works, and the conversion is a standard kinematic technique. It just needs an explicit quantitative check.\n\nAnother minor point: the reward table is dense and some weights look hand-tuned, but that's expected for this kind of framework paper. The citation pattern is honest; they cite IsaacGymEnvs, legged-gym, rsl_rl, and humanoid-gym in the acknowledgment. The code link is provided—I haven't verified the repo compiles, but the claim is checkable.\n\nWho is this for? Practitioners who want a starting point for training humanoid locomotion policies and deploying them on parallel-ankle robots. Researchers already deep in sim-to-real won't find new learning principles, but the framework and the conversion module are useful reusable pieces. It deserves a serious referee—conditional accept with a request for quantitative transfer evaluation.\n\nMy recommendation: send it to review, but require the authors to add quantitative sim-to-real comparison (tracking error, success rates) and a direct validation of the series-parallel conversion module. If the code checks out, this becomes a worthwhile community resource.","headline":"A useful open-source humanoid RL training and deployment pipeline with real hardware validation, but the 'seamless zero-shot transfer' claim rests on qualitative evidence and one under-validated kinematic conversion module.","tokens_in":9546,"tokens_out":2105,"would_cite":true,"duration_ms":20142,"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":"Booster Gym is an open-source, end-to-end reinforcement learning framework that claims to take humanoid locomotion from simulation training to zero-shot deployment on real hardware, demonstrated on the Booster T1 robot.","keywords":["humanoid robot locomotion","sim-to-real transfer","domain randomization","reinforcement learning","asymmetric actor-critic","parallel mechanism","zero-shot transfer","Isaac Gym"],"falsifier":"Run the released pre-trained policy on a Booster T1 with the series-parallel ankle conversion disabled; if the robot cannot stand or walk, the claim that the conversion module is what enables zero-shot transfer on the parallel ankle is contradicted. Also, record ten 0.5 m/s forward-walking trials in Isaac Gym and on hardware, and check whether the real joint position-velocity trajectories stay inside the envelope produced by the randomized training environments; if they leave that envelope systematically, the domain-randomization coverage claim fails.","tokens_in":8597,"feed_emoji":"🤖","tokens_out":6336,"duration_ms":62276,"temperature":0.7,"pith_summary":"Booster Gym is presented as a single open-source code framework that covers the whole humanoid-locomotion pipeline: setting up a simulation environment, training a policy with reinforcement learning, and deploying it on a physical robot without extra tuning. The authors validate it on the Booster T1 humanoid, where policies trained in Isaac Gym transfer zero-shot and produce omnidirectional walking, walking over grass, stone, soil, asphalt, concrete and tile, ascent of a 10-degree slope, and recovery after being struck by a falling 10 kg weight. The central claim is that the combination of domain randomization, a periodic gait reward, careful episode design, and a series-parallel ankle conversion module makes the sim-to-real gap small enough for direct hardware deployment. If true, this gives the robotics community a reusable starting point that removes most of the implementation burden usually associated with sim-to-real humanoid locomotion.","feed_headline":"Humanoid policies trained in simulation walk a real robot with zero tuning","feed_subtitle":"Open-source Booster Gym covers training, randomization, and parallel-ankle deployment, demonstrated on the Booster T1.","key_machinery":"The load-bearing mechanism is the combined training-and-deployment pipeline itself, held together by three parts: an asymmetric actor-critic PPO loop running in a GPU-parallel simulator; a domain-randomization envelope over robot dynamics, actuators, latency, terrain and disturbances; and a series-parallel ankle conversion module in the deployment SDK. The conversion module computes virtual serial-joint position and velocity feedback from the parallel ankle kinematics, then converts the policy's serial-joint targets into parallel-structure commands using a transposed Jacobian and the motor PD controller. This is what lets the policy ignore the closed kinematic chain during training while still controlling it on hardware. The gait-cycle observation and periodic gait reward carry the rhythmic stepping behavior, and the latency randomization matches the measured 9–12 ms command-to-feedback delay.","core_discovery":"On the paper's own terms, the discovery is that an end-to-end reinforcement learning pipeline, without teacher-student distillation or real-world fine-tuning, can be assembled into a software artifact that transfers zero-shot to a physical humanoid. The policy is trained with an asymmetric actor-critic and PPO, outputs joint position offsets at 50 Hz, and relies on a motor-level PD controller to produce torques. The actor observes noisy proprioception, the previous action, a velocity command, and a sinusoidal gait cycle, while the critic receives privileged simulation state. Domain randomization covers link mass and center of mass, joint stiffness, damping, friction, communication latency, contact properties, terrain, and external pushes, and the paper quantifies real-world latency at 9–12 ms round-trip with sub-1 ms inference to justify a 0–20 ms randomization range. A series-parallel ankle conversion module lets the policy be trained as if the ankle were a serial mechanism and then deploys on the physical parallel mechanism through a transposed Jacobian and PD control, which is the design that makes closed kinematic chains compatible with the GPU simulator.","pith_inferences":["A testable extension is to apply the same series-parallel conversion idea to other closed-chain joints, such as parallel knees or wrists; the paper only demonstrates the ankle, so this is an extrapolation.","The success of a pure randomization envelope hints that explicit adaptation mechanisms, such as a student network estimating latent dynamics, may be unnecessary for this robot class, but the paper does not run that comparison.","The community value of the released pretrained policy will depend on how easily the code retrains for a different humanoid; that portability is not tested and would be a useful follow-up benchmark.","Ablating the series-parallel conversion or the latency randomization would show which component actually closes the remaining sim-to-real gap; the paper reports no such ablation."],"forward_implications":["A new user can start from the released policy and repository rather than reimplementing environment, reward, training and deployment code, which is the practical claim behind calling Booster Gym a community resource.","The measured latency values (9–12 ms round-trip, sub-1 ms inference) give other developers concrete numbers to reproduce or tune when working on similar humanoid hardware.","Robots with parallel ankle mechanisms can train in standard GPU simulators that do not natively support closed kinematic chains, because the conversion module handles the parallel structure at deployment time.","Training on randomized terrain and disturbance conditions appears sufficient for outdoor surfaces and impact recovery without explicit terrain perception or privileged exteroception in the actor.","Cross-simulation evaluation in MuJoCo and Webots can serve as a fast validation layer that approximates real-world dynamics before hardware experiments."],"supporting_citations":[{"why":"Introduces domain randomization as a sim-to-real technique, which the framework extends to robot, actuator and environment parameters.","marker":"[5]"},{"why":"Provides the GPU-accelerated Isaac Gym simulator in which all policies are trained.","marker":"[13]"},{"why":"Supplies the massively parallel RL training approach that the framework builds on for fast policy learning.","marker":"[14]"},{"why":"An earlier humanoid zero-shot sim2real framework whose cross-simulation validation approach this paper follows.","marker":"[17]"},{"why":"Provides the asymmetric actor-critic architecture that lets the critic use privileged state while the actor sees only noisy proprioception.","marker":"[18]"},{"why":"Gives the PPO algorithm used to optimize the policy and value networks.","marker":"[19]"},{"why":"Supplies the periodic reward and gait-cycle observation that generate the bipedal walking rhythm.","marker":"[21]"}],"fun_headline_variants":["Sim-trained humanoid walks Booster T1, zero tuning","Open-source RL framework delivers zero-shot humanoid walking","No real-world fine-tuning: humanoid RL transfers instantly","Booster Gym: end-to-end RL for humanoid locomotion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Isaac Gym's simulation, once augmented by the series-parallel ankle conversion and the chosen domain-randomization ranges, is faithful enough that a policy trained on a virtual serial ankle will control the physical parallel ankle correctly without any real-world adaptation.","fun_headline_variants_meta":{"raw":{"variants":["Sim-trained humanoid walks Booster T1, zero tuning","Open-source RL framework delivers zero-shot humanoid walking","No real-world fine-tuning: humanoid RL transfers instantly","Booster Gym: end-to-end RL for humanoid locomotion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1540,"prompt_tokens":939,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":533}},"tokens_in":555,"tokens_out":601,"duration_ms":6311,"temperature":1.0,"reasoning_tokens":533,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:42:20.923268+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released pre-trained policy on a Booster T1 with the series-parallel ankle conversion disabled; if the robot cannot stand or walk, the claim that the conversion module is what enables zero-shot transfer on the parallel ankle is contradicted. Also, record ten 0.5 m/s forward-walking trials in Isaac Gym and on hardware, and check whether the real joint position-velocity trajectories stay inside the envelope produced by the randomized training environments; if they leave that envelope systematically, the domain-randomization coverage claim fails.","supporting_citations":[{"cited_title":"Sim-to-real learning of all common bipedal gaits via periodic reward composition,","cited_arxiv_id":null,"evidence_quote":"Supplies the periodic reward and gait-cycle observation that generate the bipedal walking rhythm."}],"review_version":2}