{"id":"f76dc1d8-57c7-447a-9fbd-67ce25caf382","arxiv_id":"2505.09979","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-level learning controller lets a quadruped robot perform dog-like gaits and complete an agility course with depth-only perception at an average 1.1 m/s, peaking at 3.2 m/s while hurdling.","lead":"Researchers trained a quadrupedal robot to switch between natural dog-like gaits (walk, trot, canter, jump) and used those behaviors to complete an obstacle course at 1.1 m/s average speed using only a depth camera. The work combines imitation learning, simulator tuning, and a task-level controller, and it introduces a benchmark for robot agility.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's real-world success rate (>78%) is reported without trial counts, error bars, or a trial definition; the number of real runs is the least-secure condition for the headline claim.","rationale":"The paper's contribution is an engineering demonstration: a Go2 with depth perception completes a randomized dog-agility course. The load-bearing condition is that the robot reliably completes the course, not once but across randomized configurations; otherwise 'success rate exceeds 78%' is meaningless. I considered the sim-to-real fidelity concern raised by the reader: EASI optimizes only motor PD parameters (Supplementary S2) and Fig.6c-d is qualitative. That is a real concern, but it is secondary because the paper's direct evidence for transfer is the real-world completion itself; if the trial log showed, say, 19/20 runs completed, the EASI mechanism would be adequately supported even if the frequency spectra are informal. Conversely, a cherry-picked run would sink the claim even if EASI were perfect. I also checked for internal inconsistencies: Section 2.3 claims canter speed tracking up to 3.5 m/s while Table S2 caps canter vx at 2.0 m/s, and the claimed 3.2 m/s peak exceeds that cap; this should be clarified but does not by itself undermine the completion claim. The reader's rationale already notes missing trial counts, but their formal weakest_assumption was EASI fidelity, so my agreement is partial. The appropriate verdict remains CONDITIONAL: the architecture is plausible and the videos are evidence of at least one successful run, but the quantitative success-rate claim should be warranted by released per-trial data before being taken at face value.","tokens_in":19773,"tokens_out":9515,"duration_ms":96663,"concrete_test":"Ask the authors for the raw trial log of the full agility challenge: for each run, the sampled obstacle order/positions, pass/fail, completion time, and whether any operator intervention or reset occurred. Then compute the exact two-sided 95% Clopper-Pearson interval for the success probability. Also request the equivalent per-trial data for Fig.2d-e (single-obstacle and height sweeps) and a statement confirming that the '>78%' figure in Section 2.1 is from physical trials rather than the Fig.5b simulation. If N<10, the data cannot distinguish a 78% success rate from a 50% one, and the headline should be downgraded to a qualitative demonstration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the integrated controller 'successfully complete[s] the quadrupedal agility challenge' with a success rate that 'exceeds 78%' (Section 2.1). This is a statistical statement about repeated randomized trials, but the paper supplies none of the statistics: no number of real-world runs, no per-trial outcomes, no obstacle-configuration sampling protocol, no definition of failure (knock-down, fall, timeout, operator reset). Fig.2f shows a single point without error bars, and the 1.1 m/s average comes from one 26 s run. The provenance of the 78% is also ambiguous: Section 2.4 reports a '~78%' success rate for the simulated noise-robustness experiment, and Section 2.1 uses the same threshold for real-world trials, so the reader cannot tell whether the real number is an independent measurement or a transferred simulation value. Without N, 'exceeds 78%' is compatible with 4/5 as well as 20/25, and one failed trial would move 4/5 to 60%. The central claim therefore rests on an unquantified empirical base, and the paper's own no-code/no-data policy makes it impossible to audit. This is a load-bearing gap, not just a stylistic preference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an integrated controller for a Unitree Go2 quadruped, combining a Basic Behavior Controller (BBC) trained by a semi-supervised variant of InfoGAIL from dog motion capture data, a Task-Specific Controller (TSC) trained by privileged learning with depth images, and an Evolutionary Adversarial Simulator Identification (EASI) procedure for sim-to-real transfer. The authors report that the robot completes a randomized quadrupedal agility course at an average speed of 1.1 m/s, reaches 3.2 m/s during hurdling, achieves a success rate exceeding 78% on the full course, and exhibits five natural gait modes with continuous style variation, with ablations for the BBC, TSC, and EASI components.","tokens_in":20070,"tokens_out":7359,"duration_ms":77807,"significance":"If the empirical claims hold, this is a meaningful advance at the intersection of imitation learning, hierarchical control, and sim-to-real transfer: it demonstrates a single integrated controller that switches among diverse natural behaviors in real time using depth perception alone. The paper's explicit strengths include thorough ablations of the BBC (NMI/ENT metrics and t-SNE visualizations), comparisons against a strong parkour baseline, real-robot deployment, and a sim-to-sim validation of the EASI parameter search. The reported training times and data requirements are also concrete and reproducible in spirit. However, the central empirical claims are currently supported by a single 26 s run and by success-rate values with no trial counts, error bars, or failure definitions, and one of the formal objectives in the EASI derivation appears mis-specified. The overall approach is credible and potentially important, but the evidence base needs substantial strengthening before the headline claims can be accepted.","major_comments":[{"comment":"The headline claim that the robot 'successfully completes the quadrupedal agility challenge' with a success rate that 'exceeds 78%' is not statistically supported. The full-course result is presented as a single 26 s run (Fig. 2b), with no number of randomized trials, per-trial outcomes, confidence interval, or definition of failure (e.g., whether obstacle knock-down, falls, timeouts, or operator resets count as failures). In addition, Section 2.4 reports a simulated noise-robustness success rate of approximately 78% for the TSC, so it is unclear whether the real-world number is an independent measurement or a transferred simulation value. The authors should report trial counts and the full evaluation protocol, and present all success rates with error bars.","section":"Section 2.1 and Section 2.4"},{"comment":"Equation (13), as written, does not define the adversarial game described in the text. In a standard LSGAN formulation, the discriminator minimizes E_fake[(D+1)^2] + E_real[(D-1)^2] while the generator should minimize E_fake[(D-1)^2] (i.e., push fake transitions toward the real label). The paper instead states max_Ξ min_~D of that full expression, which would make the generator maximize the fake term E[(D+1)^2] and would drive the outer objective toward a degenerate solution because the inner minimization over ~D can set D(fake)=-1 independent of Ξ. This appears to be more than a notational slip: the sign and the min-max order need to be corrected or clarified, and the actual objective used in Algorithm S2 should be stated explicitly.","section":"Section 4.4, Eq. (13)"},{"comment":"The NMI/ENT comparison of behavioral diversity depends on a 'pre-trained behavior classifier' whose training labels, architecture, and accuracy are not specified. If this classifier is trained on the same five human-defined categories that the latent skill variable c is designed to encode, then the reported NMI/ENT values primarily demonstrate that the policy can be controlled by the latent code, rather than that the method discovers behavior modes from raw data. The paper should document the classifier and provide additional, independent validation of the naturalness and separability of the generated behaviors (e.g., human evaluation or ground-truth label metrics), so that the diversity claim is not circular.","section":"Section 2.3"},{"comment":"The evidence for the claim that the BBC transfers to the real world 'with minimal performance loss' after EASI fine-tuning is qualitative, consisting of joint frequency spectra and trajectories in Fig. 6c-d. The sim-to-sim experiment in Fig. 6a shows convergence to true parameters, but the sim-to-real experiment in Fig. 6b lacks quantitative transfer metrics such as success rate, tracking error, or task reward measured on the real robot. Since EASI is a core component of the proposed pipeline, the paper should provide quantitative sim-to-real transfer results with trial counts rather than relying on visual similarity of signals.","section":"Section 2.5 and Section 4.4"}],"minor_comments":[{"comment":"Please clearly define the success/failure criteria for the agility challenge before the first use of the 'success rate' phrase, and state explicitly whether the 26 s completion time and the 1.1 m/s average speed come from one run or from an average over multiple runs.","section":"Section 2.1"},{"comment":"The manuscript contains several typos: 'IssacGym' in Section 4.1 should be 'IsaacGym'; 'the the sim-to-real gap' in Section 1 has a duplicated article; and 'The results shows' in Section 2.3 should be 'The results show.'","section":"General"},{"comment":"The text uses both 'exceeds 78%' and 'approximately 78%' for the success-rate claim; please disambiguate the real-world and simulated measurements and avoid carrying the same numerical threshold across different experiments.","section":"Section 2.1 and Section 2.4"},{"comment":"In Equation (3), the notation d_E^L(o_I|y) is not defined before use; please specify the conditional distribution of the observation window given the label for the labeled expert demonstrations.","section":"Section 4.3, Eq. (3)"},{"comment":"The paper should include a data and code availability statement, or at least a permanent repository link, so that the reported trial counts and learned policies can be audited by the community.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main risk to publication is the lack of trial counts and error bars for the headline real-world success rate. The similarity between the Section 2.1 value ('exceeds 78%') and the Section 2.4 simulated value ('approximately 78%') raises the possibility that the same number was carried over rather than independently measured; please ask the authors to provide the raw trial log. The formal issue in Eq. (13) is also concerning and should be resolved before acceptance, as it affects the written derivation of a core method component."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real system paper with a genuinely new integrated result—a depth-only controller that switches between learned dog-like gaits and completes a six-obstacle agility course on a Unitree Go2—and its main quantitative claim (real-world success rate >78%) is under-documented. The paper deserves a serious referee, but the headline number needs more than one point.\n\nWhat's new: the BBC's semi-supervised skill+style decomposition is a modest but real extension of the authors' ESS-InfoGAIL, and the full BBC+TSC stack, including the EASI-identified simulator and BYOL-trained depth encoder, is demonstrated on hardware. The benchmark itself is useful. The ablations are thorough: NMI/ENT, command tracking, comparisons against GAIL/InfoGAIL/ACGAIL, teacher/student ablations, and the EASI sim-to-sim and sim-to-real evaluations. The math in the mutual-information objectives is standard variational bounds, correctly derived; no red flags there.\n\nSoft spots, in order of importance. (1) The 'success rate exceeds 78%' is not backed by statistics anywhere: no trial count, no per-trial outcomes, no failure definition, no error bars. Fig.2f shows one point. Worse, Section 2.4 reports ~78% in a simulated noise-robustness experiment, so the real-world number could be a transferred simulation value or an independent measurement—the text doesn't let you tell. This is a load-bearing gap for the paper's central claim. (2) The sim-to-real transfer claim rests on qualitative frequency spectra and trajectories (Fig.6); the 80-second real dataset is small, and no quantitative real-world tracking or success metrics per condition are given. (3) The latent skill modes are human-labeled, so the 'diverse behavior discovery' is really controllable generation of a pre-defined taxonomy; that's fine, but the paper should say it plainly. (4) No code or data released, which makes the reproducibility ceiling lower.\n\nWho is this for: people working on legged locomotion, imitation learning, and sim-to-real. A serious referee should engage with it, mainly to chase the real-world trial statistics and the exact protocol for the agility challenge. Recommendation: send to peer review, conditional on the authors providing N, trial-level results, and a clear separation between simulated and real success rates. If they can't, the 78% claim should be softened.","headline":"A real depth-only quadruped controller that switches learned gaits and completes an agility course on hardware; the headline success rate is real but statistically under-documented.","tokens_in":20589,"tokens_out":1759,"would_cite":true,"duration_ms":17942,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-layer controller that imitates five real dog gaits and reads only depth images lets a quadrupedal robot run a randomized six-obstacle agility course at 1.1 m/s, reaching 3.2 m/s over hurdles.","keywords":["quadrupedal robotics","imitation learning","sim-to-real transfer","adversarial imitation learning","privileged learning","depth perception","locomotion","dog agility"],"falsifier":"Run the full randomized agility challenge twenty times with the same controller and count per-trial successes: the claimed success rate above 78% predicts no more than four failures, and a systematic failure on any single obstacle would expose a gap in the behavior library rather than in perception. Separately, command motions outside the 80-second calibration set (sustained 3 m/s trotting, or sharp turning on a different floor surface) and compare real joint trajectories with simulator rollouts under EASI's identified parameters: divergence beyond the calibration-time spread would refute the minimal-loss transfer claim.","tokens_in":19572,"feed_emoji":"🐕","tokens_out":10274,"duration_ms":97871,"temperature":0.7,"pith_summary":"This paper claims that a quadrupedal robot can achieve dog-like agility by splitting control into two learned layers: a Basic Behavior Controller that imitates diverse real-dog gaits, and a Task-Specific Controller that decides, from depth-camera images alone, which gait to use and how fast to go. On a Unitree Go2 robot, the authors report the first autonomous completion of a randomized six-obstacle agility course — A-frame, bar-jump, poles, seesaw, tire-jump, and tunnel — at an average speed of 1.1 m/s with a success rate above 78%, plus a peak speed of 3.2 m/s in a hurdling task. The decisive step is transfer: they tune simulator physics with about 80 seconds of real-world data, fine-tune the behavior controller inside that improved simulator, and argue the resulting behaviors transfer with minimal performance loss. If true, the paper shows that broad, natural, animal-like behavior libraries can be built from motion-capture data and shipped to real robots with cheap sensing, opening agility-style autonomy to commodity hardware.","feed_headline":"Robot clears dog agility course at 1.1 m/s on depth vision alone","feed_subtitle":"Five learned dog gaits plus an 80-second real-world tune-up carry a Go2's agility from simulation to reality.","key_machinery":"The argument rests on three components. The BBC is trained with semi-supervised InfoGAIL: a least-squares discriminator (the AMP-style reward) scores how closely transitions match real dog motion, while a predictor maximizes mutual information between latent variables and transitions, so the policy learns to produce specific gaits on command — the discrete skill latent $c$ chooses the mode, the continuous shift latent $\\epsilon$ varies style within a mode, and RIM lets the latent distribution rebalance toward the true, imbalanced data. The TSC is a student policy trained by privileged learning: a teacher trained with Hybrid-PPO in a hybrid action space sees scandots, waypoint yaws, and obstacle type, and the student, which sees only depth images through a GRU encoder, learns to imitate it while a BYOL contrastive loss and probabilistic augmentations (white noise, blur, cropping) make the depth features robust. The third component, EASI, turns simulator identification into an adversarial search: an evolutionary strategy proposes distributions over joint PD gains, a neural discriminator scores transitions by realism, and the loop converges in under ten minutes on 80 seconds of real data, after which the BBC is fine-tuned in the enhanced simulator.","core_discovery":"On its own terms, the paper's central discovery is that quadrupedal agility decomposes into a reusable behavior library and a task coordinator, and that both can be learned and transferred. The Basic Behavior Controller is a single policy, trained by a semi-supervised variant of InfoGAIL, that reproduces five dog gaits — walk, pace, trot, canter, jump — from raw motion capture data with less than 5% of it labeled: a discrete latent variable selects the gait, a continuous latent variable selects style, and a Regularized Information Maximization term lets the policy learn from imbalanced unlabeled data. The Task-Specific Controller is a student policy trained by privileged learning to imitate a teacher that sees scandots, waypoint angles, and obstacle types; the student sees only a depth image, with a BYOL self-supervised loss and random augmentations shielding it from real-world image noise. The simulator is then brought closer to reality by EASI, an evolutionary search over motor parameters scored by a neural discriminator, using 80 seconds of real-world data and under ten minutes of GPU time. With these pieces, the robot switches between trot, canter, and jump by itself, jumps onto a 30 cm box, and clears hurdles at 3.2 m/s — evidence, the paper argues, that a substantial step toward animal-like agility has been made.","pith_inferences":["Because EASI accepts any controller for collecting calibration trajectories, the 80-second real-world step could be automated and repeated over the robot's lifetime, letting the simulator track actuator wear, battery voltage, or payload changes — a maintenance loop the paper does not explore.","The BBC's controllability via two latent variables suggests the same library could be driven by other high-level interfaces, including voice- or gesture-based human commands, which the paper lists as future work; a low-cost teleoperation test would be a direct way to verify this.","The student's reliance on a short depth history invites a concrete test of whether the reported success rate degrades with a narrower camera field of view or longer obstacle runs; if it does, the limiting factor is perception range rather than the gait library, redirecting effort toward wider-angle or multimodal sensing.","Nothing restricts the behavior library to dogs: retraining the BBC on motion capture from other quadrupeds or on synthetic gaits could probe whether the agility gains come from the specific five gaits themselves or from the semi-supervised training objective that disentangles them."],"forward_implications":["A quadruped with a single depth camera can navigate a randomized course of six visually distinct obstacles without global localization, so agility does not require external tracking or motion-capture infrastructure.","Because the BBC is task-agnostic, the same five-gait library can serve new downstream tasks by training only a new TSC; box jumping, hurdling, and human-collaborative routing are demonstrated instances of this reuse.","The reported success rate above 78% at 1.1 m/s on a commercial robot, against the baseline parkour controller, sets a concrete benchmark — the paper's own agility challenge — that other groups can reproduce and beat.","EASI's use of only 80 seconds of real data and under ten minutes of GPU search suggests that simulator tuning can replace hand-set domain randomization ranges as the standard sim-to-real step for legged robots.","The robot clears bar-jumps up to 0.35 m and tire-jumps up to 0.7 m by switching autonomously from canter to jump using depth alone, which the paper shows beats both the parkour baseline and a human teleoperator on the same course."],"supporting_citations":[{"why":"Supplies the raw dog motion capture data whose five gait modes the BBC must reproduce.","marker":"[57]"},{"why":"The InfoGAIL algorithm this work extends with semi-supervised labels and a RIM objective.","marker":"[34]"},{"why":"The prior semi-supervised InfoGAIL formulation for imbalanced demonstrations that anchors the BBC training objective.","marker":"[44]"},{"why":"The EASI method itself, which tunes simulator joint-PD parameters against 80 seconds of real data.","marker":"[41]"},{"why":"The baseline parkour controller used for comparison in obstacle-course experiments.","marker":"[21]"},{"why":"The AMP-style least-squares discriminator and imitation reward that give the BBC its dog-like motion style.","marker":"[33]"},{"why":"BYOL, the self-supervised contrastive objective that makes the depth encoder robust to real-world image noise.","marker":"[48]"},{"why":"Hybrid-PPO, the hybrid action-space algorithm used to train the TSC teacher that commands both gaits and speeds.","marker":"[47]"}],"fun_headline_variants":["Robot dog clears hurdles at 3.2 m/s using five learned gaits","Quadruped robot learns dog-like gaits, achieves 3.2 m/s hurdle speed","Sim-to-real robot learns five dog gaits, jumps 30 cm box","Robot dog's five learned gaits enable 3.2 m/s hurdle runs","From raw mocap to real robot: five dog gaits and 3.2 m/s hurdle speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results depend on the simulator, after EASI tuning on about 80 seconds of real-world data, matching the real robot's dynamics closely enough that the fine-tuned BBC transfers with minimal performance loss; the paper's evidence for this match is qualitative — joint frequency spectra and trajectories — rather than a measured success-rate comparison in simulation versus reality.","fun_headline_variants_meta":{"raw":{"variants":["Robot dog clears hurdles at 3.2 m/s using five learned gaits","Quadruped robot learns dog-like gaits, achieves 3.2 m/s hurdle speed","Sim-to-real robot learns five dog gaits, jumps 30 cm box","Robot dog's five learned gaits enable 3.2 m/s hurdle runs","From raw mocap to real robot: five dog gaits and 3.2 m/s hurdle speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001546,"raw_usage":{"total_tokens":6237,"prompt_tokens":1054,"completion_tokens":5183,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":5070}},"tokens_in":670,"tokens_out":5183,"duration_ms":32265,"temperature":1.0,"reasoning_tokens":5070,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:19:31.816459+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full randomized agility challenge twenty times with the same controller and count per-trial successes: the claimed success rate above 78% predicts no more than four failures, and a systematic failure on any single obstacle would expose a gap in the behavior library rather than in perception. Separately, command motions outside the 80-second calibration set (sustained 3 m/s trotting, or sharp turning on a different floor surface) and compare real joint trajectories with simulator rollouts under EASI's identified parameters: divergence beyond the calibration-time spread would refute the minimal-loss transfer claim.","supporting_citations":[{"cited_title":"& Saito, J","cited_arxiv_id":null,"evidence_quote":"Supplies the raw dog motion capture data whose five gait modes the BBC must reproduce."},{"cited_title":"& Ermon, S","cited_arxiv_id":null,"evidence_quote":"The InfoGAIL algorithm this work extends with semi-supervised labels and a RIM objective."},{"cited_title":"Ess-infogail: Semi-supervised imitation learning from imbalanced demonstrations, Vol","cited_arxiv_id":null,"evidence_quote":"The prior semi-supervised InfoGAIL formulation for imbalanced demonstrations that anchors the BBC training objective."},{"cited_title":"& Chen, C.Easi: Evolutionary adversarial simulator identification for sim-to-real transfer, Vol","cited_arxiv_id":null,"evidence_quote":"The EASI method itself, which tunes simulator joint-PD parameters against 80 seconds of real data."},{"cited_title":"& Pathak, D.Extreme parkour with legged robots, 11443–11450 (IEEE, 2024)","cited_arxiv_id":null,"evidence_quote":"The baseline parkour controller used for comparison in obstacle-course experiments."},{"cited_title":"B., Ma, Z., Abbeel, P., Levine, S","cited_arxiv_id":null,"evidence_quote":"The AMP-style least-squares discriminator and imitation reward that give the BBC its dog-like motion style."},{"cited_title":"& Yu, Y.Hybrid actor-critic reinforcement learning in parameterized action space, 2279–2285 (IJCAI,2019)","cited_arxiv_id":null,"evidence_quote":"Hybrid-PPO, the hybrid action-space algorithm used to train the TSC teacher that commands both gaits and speeds."}],"review_version":1}