{"id":"33bf1426-668c-40d2-8fc3-7131fabcc808","arxiv_id":"2505.16042","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A single dynamics-conditioned RL policy transfers zero-shot across quadrupeds from 12 kg to 50 kg, and diverse reference robots during training clearly improve tracking.","lead":"This paper trains one reinforcement-learning controller that can walk on very different quadruped robots by feeding it a learned signal about each robot's dynamics. The main practical finding is that training on several robot families instead of one cuts velocity-tracking error by up to 30% and lets the same policy drive a 50 kg ANYmal C with no fine-tuning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'zero-shot transfer' claim may be in-distribution: ANYmal C is a training reference model (ID 5) whose randomized variants populate the training set, so the hardware results do not establish generalization to a truly unseen quadruped.","rationale":"Good-faith reading: the paper's contribution is not that the GRU-based PAL beats MorAL—the paper explicitly reports the opposite on ANYmal C—but that a single policy trained over diverse reference models can zero-shot transfer to a range of quadrupeds and that reference diversity itself improves tracking. I see the reported negative result as honest, and the three-seed learning curves in Fig. 3 are a plus. The load-bearing issue is the operational definition of 'zero-shot.' Because ANYmal C is one of the reference models (ID 5), and the training set consists of 50 randomized variants per reference model drawn from intervals that include the nominal robot's parameters, the hardware evaluation on ANYmal C does not test a robot outside the training distribution. The authors' caveat that 'the specific quadrupeds were never included in the training sets' is technically true but weak: the randomized variants are drawn around the same base model, and the admission filter (2 s standing) accepts the nominal geometry. The only ID set that leaves out ID 5 ({1,4}) shows substantially worse RMSE on ANYmal C, which is consistent with the hypothesis that the apparent success is in-distribution interpolation. This does not invalidate the reference-diversity finding, but it does mean the central zero-shot claim is not yet supported. A leave-one-reference-out experiment, or a quantitative coverage check, would settle it. Since this is addressable with additional evaluation rather than a fundamental flaw, the conditional verdict stands unchanged, with the condition sharpened to require a true out-of-distribution target or a demonstrated coverage guarantee.","tokens_in":12430,"tokens_out":7838,"duration_ms":69086,"concrete_test":"Re-train the pipeline with ANYmal C's reference model held out (ID set {1,2,4}) and evaluate on the real ANYmal C under the same walking protocol as Table V. If RMSE-X and RMSE-theta are statistically unchanged from the {1,2,4,5} policy (e.g., within one standard error or <10% relative), the zero-shot claim survives; if the gap reproduces the observed 18-30% degradation, ANYmal C is an in-distribution result. As a secondary check, compute the normalized parameter-space distance (e.g., Mahalanobis distance over Table I ranges) from the nominal ANYmal C to the admitted training samples; if the nominal robot is inside the high-density support of ID-5 variants, it cannot be called an unseen platform.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B.1 states that 'We utilize up to four simplified reference models based on existing quadrupeds to generate a set of 50 robot configurations per model,' and Table I gives per-ID sampling ranges that include ID 5 = ANYmal C. The test protocol in Section IV.B adds that 'the specific quadrupeds were never included in any of the training sets, but instead used as a base model in which we change the sampled kinematic and dynamic parameters.' This distinction does not make the test robot unseen: the nominal ANYmal C lies inside or near the support of the ID-5 variant distribution (e.g., mbase is sampled from U(18,50) kg, so the 50 kg ANYmal C is at the upper edge), and the 2 s stand collision-free admission filter would admit the nominal robot. Policies trained with the ID set {1,2,4,5} are therefore evaluated on a robot whose nominal configuration is within the training distribution. The hardware rows that do exclude ID 5 (PAL/MorAL with {1,4}) are meaningfully worse (PAL RMSE-X 0.2019 vs 0.1652; MorAL RMSE-X 0.1243 vs 0.0885), which is consistent with the hypothesis that the apparent 'zero-shot' success on ANYmal C is interpolation over training reference models rather than extrapolation to a new morphology. No coverage metric or leave-one-reference-out experiment is provided to show otherwise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Platform Adaptive Locomotion (PAL), a single reinforcement-learned controller for quadrupedal robots with diverse morphologies and dynamics. The policy is trained on procedurally randomized variants of several reference quadruped models (A1, Aliengo, ANYmal B, ANYmal C) and is conditioned on either a GRU-based latent dynamics embedding (DIM) or an explicit morphology estimator (MorAL-style). The authors compare PAL against MorAL and GenLoco baselines under different sets of training reference models, measuring simulation success rates under perturbations and hardware velocity-tracking performance on ANYmal C. The reported results show that training on a broader set of reference models improves robustness and tracking, that PAL achieves competitive simulation robustness on the smaller A1, and that the MorAL baseline achieves lower tracking error on ANYmal C hardware. The paper also documents a negative result: PAL does not outperform MorAL on the ANYmal C hardware task.","tokens_in":12793,"tokens_out":5850,"duration_ms":50310,"significance":"If the central claims were fully substantiated, the paper would make a useful contribution: a reference-free policy that transfers across a 12–50 kg range of quadrupeds without per-robot retraining, with explicit ablations over the choice of training reference models. The paper is commendably transparent about its negative hardware result for PAL versus MorAL, and it provides reproducible-looking training details and multi-seed learning curves in Figure 3. However, the significance is currently limited by the fact that the 'zero-shot transfer' claim is not actually demonstrated: the two robots evaluated in detail (A1 and ANYmal C) are themselves training reference models whose nominal parameters lie inside the sampled distributions, and the hardware comparison lacks trial counts, error bars, and ground-truth velocity references.","major_comments":[{"comment":"The claim of zero-shot transfer to ANYmal C and A1 is not supported by the evidence, because both robots are used as reference models for generating the training distribution. Table I samples the base mass of ID 5 (ANYmal C) from U(18,50) kg, so the nominal 50 kg ANYmal C lies at the upper edge of the trained mass range, and ID 1 (A1) has its nominal mass inside U(2,28) kg. Section IV.A states that 'the specific quadrupeds were never included in any of the training sets, but instead used as a base model in which we change the sampled kinematic and dynamic parameters'; this is true only in the sense that the exact nominal parameter vector is not sampled, but the nominal configuration is still within or near the support of the training distribution. A leave-one-reference-out experiment (e.g., train on IDs {1,2,4} and test on nominal ANYmal C, or evaluate on ID 3, which is never used in the multi-ID training sets) is needed to substantiate the word 'zero-shot'. Without such an experiment, the hardware and simulation results are better described as interpolation over training reference models.","section":"IV.A, IV.B, Table I"},{"comment":"The hardware transfer to ANYmal C is not zero-shot with respect to actuation dynamics. Section III-B.4 states that the actuator networks which randomly replace Equation (3) during training were trained on data collected on ANYmal B and C hardware. Since the hardware deployment target is ANYmal C, the policy's training distribution already includes the target robot's actuator characteristics. This should be acknowledged as a limitation, or the actuator networks should be retrained without ANYmal C data to preserve the zero-shot claim.","section":"III-B.4, IV.C"},{"comment":"The hardware RMSE tables contain no trial counts, per-run variance, or significance tests. The reported differences (e.g., PAL RMSE-X 0.1652 vs 0.2019, or MorAL RMSE-X 0.0885 vs 0.1339) may be within run-to-run variability on hardware, especially given the small number of conditions. Additionally, Table V compares the target velocity command with the learned velocity estimator, not with ground-truth base velocity, so estimator errors are conflated with tracking errors. The paper should report the number of trials, error bars, and ideally an independent motion-capture or onboard-state-estimator cross-check for at least one condition before claiming percentage improvements.","section":"IV.D, Tables V and VI"},{"comment":"The MorAL baseline is modified relative to the original method: the height map is removed and the action history length differs from the original MorAL policy. The paper does not provide an ablation quantifying the effect of these changes on MorAL's performance. Since the central hardware conclusion is that MorAL outperforms PAL, it should be shown that these modifications do not bias the comparison in either direction.","section":"IV.A, III.E"}],"minor_comments":[{"comment":"The claim of 'reducing the velocity tracking error by up to 30%' is ambiguous about which controller and which baseline are being compared. In Table V, the 30+% reduction appears for MorAL (RMSE-X 0.0885 vs 0.1339), while PAL shows about 25% reduction (0.1652 vs 0.2193); the wording should be clarified.","section":"Abstract, V"},{"comment":"The text says 'The same early termination criteria defined in Section III-B are applied,' but Section III-B describes the robot admission filter, not early termination criteria; the early termination penalty is described later in Section III-G. Please correct the cross-reference.","section":"IV.B"},{"comment":"The air-time reward expression is malformed: 'rair,i = −3 · ( if ||cdes||2 = 0 : −Ti,swing else: Ti,swing − 0.5' is missing a closing parenthesis and the intended sign convention is unclear.","section":"Table III"},{"comment":"The black 'Min/Max' training ranges for the horizontal force perturbation are not defined in Table I or in the text; please provide the corresponding sampling ranges for each perturbation type.","section":"Figure 4"},{"comment":"The sentence 'Compared to the MorAL policy of Luo et al. [14], we only use the state history of the past two control steps instead of the past five' is ambiguous: it is unclear whether this change applies to the PAL policy only or to both PAL and the implemented MorAL baseline. Please clarify.","section":"III.E"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with a useful negative result, but the headline generalization claim needs substantial re-framing or additional experiments. The evaluation robots are training reference models, and the hardware results lack statistical grounding. I think major revision is appropriate: leave-one-reference-out experiments or a genuinely novel test robot, plus trial counts and error bars, would make the contribution publishable. The fit with the journal is reasonable. The authors should also be encouraged to release code and policy checkpoints; the paper already gives many training details, which is a strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper has a real result, and it is honestly reported. Training a single locomotion policy on a diverse set of quadruped reference models improves robustness and velocity tracking on hardware, and the comparison between a GRU-based dynamics encoder (PAL) and a morphology-estimator baseline (MorAL) is clean. The authors also state plainly that their method does not beat MorAL on the ANYmal C tracking task. That honesty is rare and valuable.\n\nThe main contribution is the systematic study of reference-set composition: {1}, {1,4}, and {1,2,4,5} show a clear trend, with the largest set giving up to 30% lower tracking error. The finding that explicit morphology estimation beats temporal dynamics encoding on the heavier robot is plausible and backed by the success-rate curves. The hardware deployment on a 50 kg ANYmal C with a single policy is a genuine engineering achievement, and the latency modeling that made sim-to-real work is a useful detail.\n\nThe soft spots are real but addressable. The 'zero-shot transfer' language is stronger than the evidence. ANYmal C is one of the base reference models (ID 5) used to generate training variants, and the nominal robot sits at the edge of the sampled mass distribution. The hardware rows that exclude ID 5 are noticeably worse, which is consistent with interpolation over training references rather than extrapolation to a genuinely unseen platform. A leave-one-reference-out experiment would settle this, and the absence of one is the biggest gap. The hardware tables also lack trial counts and error bars, and the velocity RMSE uses the learned estimator rather than ground truth, so the numbers should be treated with some caution. No code or data is released, which limits reproducibility of the procedural generation details.\n\nFor someone working on multi-embodiment locomotion or sim-to-real transfer, this paper is worth reading and citing. It deserves a serious referee, but the authors should be pushed to tighten the generalization claim and, ideally, add the leave-one-out experiment. The core empirical message about reference-set diversity is solid even if the packaging overshoots.","headline":"Solid empirical study of reference-set diversity, but the zero-shot transfer claim is weaker than advertised because the test robot is one of the training reference models.","tokens_in":13351,"tokens_out":2767,"would_cite":true,"duration_ms":25480,"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":"A single reinforcement-learning policy can drive quadrupeds from 12 kilograms to 50 kilograms with no per-robot retraining.","keywords":["quadrupedal locomotion","reinforcement learning","zero-shot transfer","dynamics conditioning","procedural robot generation","sim-to-real transfer","platform-adaptive control","velocity tracking"],"falsifier":"Train the same two architectures on a reference set that leaves out ANYmal C-like body masses and leg lengths, then deploy on ANYmal C and measure heading velocity RMSE. If the error stays near the full-reference values of 0.1652 for PAL and 0.0885 for MorAL instead of degrading toward the single-reference values of 0.2193 and 0.1339, the paper's claim that diverse reference modelling drives the improvement would be wrong; the expected contrast is given in the paper's own Table V.","tokens_in":12273,"feed_emoji":"🤖","tokens_out":6935,"duration_ms":55338,"temperature":0.7,"pith_summary":"The paper tries to establish that one reinforcement-learning policy can control quadrupeds of very different sizes and dynamics without per-robot retraining, by conditioning every action on a latent estimate of the robot's current dynamics. The policy is trained in simulation on procedurally generated variants of four reference quadrupeds and then transferred zero-shot to unseen simulated robots and to a real 50 kg ANYmal C. The authors report that the set of reference models chosen during training is itself a performance lever: including diverse morphologies reduces velocity command tracking error by up to 30% compared with training on a single reference model. They also find that a morphology-aware estimator of robot properties tracks velocity commands more accurately on ANYmal C hardware than their GRU-based temporal dynamics encoder, even though both approaches transfer successfully.","feed_headline":"One policy walks quadrupeds from 12 kg to 50 kg","feed_subtitle":"Dynamics-conditioned training cuts velocity tracking error by up to 30 percent on ANYmal C hardware.","key_machinery":"The load-bearing mechanism is the dynamics-conditioned observation loop: the GRU-based DIM condenses recent base and joint state history into a latent vector of robot dynamics, and the CM policy receives that vector as part of its observation when outputting joint position targets. Around this loop sits procedural robot generation: each reference model yields 50 randomised variants drawn from the parameter ranges in Table I, admitted only if the robot can stand collision-free for 2 seconds in the simulator. Actuator command latency modelling and actuator networks trained on real ANYmal data are also injected during training. Together these parts make the policy platform-adaptive without motion priors: the latent embedding lets one set of weights respond appropriately across robots, and the generation ranges determine whether that embedding covers the target platform.","core_discovery":"Platform Adaptive Locomotion (PAL) is a single reference-free locomotion policy consisting of a Dynamics Inference Module (DIM), a gated recurrent unit that reads recent base and joint history and emits a latent vector of the robot's temporally local dynamics, and a Control Module (CM) that maps proprioceptive state plus that latent vector to twelve desired joint positions tracked by PD controllers. Training happens on procedurally generated quadrupeds obtained by randomising the kinematic and dynamic properties of reference models spanning the A1, Aliengo, ANYmal B and ANYmal C. The central discovery is that conditioning on dynamics enables zero-shot transfer across quadrupeds from 12 kg to 50 kg, and that the diversity of reference models used during training is a decisive ingredient: the full reference set reduces hardware velocity tracking error by up to 30% relative to single-reference training, and the paper traces this to exposure to realistic mass distributions. On the ANYmal C hardware, the morphology-estimating MorAL architecture proved the more accurate controller, so the paper's contribution is evidence about which design choices matter rather than a claim that PAL beats every baseline.","pith_inferences":["If the trend holds, future universal legged controllers could be specified by a parameter distribution rather than a robot identity, so a new platform would be added by checking that its dynamics fall inside the trained range.","The two conditioning strategies appear complementary, since the velocity-estimator results are similar across architectures; combining the GRU dynamics encoder with a morphology-property estimator is a natural next test.","The hand-picked randomisation ranges are themselves the real hypothesis, so a formal coverage test over the target robot's parameter space would predict when zero-shot transfer succeeds and when it fails.","The 30% improvement from reference diversity suggests that scaling a single reference model, as earlier work did, under-samples the mass and inertia regimes that matter for heavier robots."],"forward_implications":["A single trained policy can be deployed across quadrupeds from the 12 kg A1 to the 50 kg ANYmal C without per-robot retraining, including zero-shot transfer in simulation and on hardware.","Selecting multiple diverse reference quadrupeds for training reduces hardware velocity tracking error by up to 30% compared with a single-reference baseline, so reference-set choice is a controllable design parameter.","Morphology-aware conditioning yields lower velocity tracking error on ANYmal C than pure temporal dynamics encoding, indicating that architecture still matters once dynamics conditioning is in place.","Explicit modelling of actuation latency was required to move from simulation to the real ANYmal C, confirming that actuator response is one of the main sim-to-real gaps."],"supporting_citations":[{"why":"Supplies the MorAL baseline and concurrent morphology-estimator architecture that PAL is compared against and that tracks velocity more accurately on ANYmal C.","marker":"[14]"},{"why":"Supplies the GenLoco baseline whose performance degrades for quadrupeds above about 20 kg, motivating dynamics conditioning.","marker":"[15]"},{"why":"Motivates the GRU-based Dynamics Inference Module used to embed robot dynamics from state history.","marker":"[17]"},{"why":"Provides the actuator-network replacement for PD control and the explicit command-latency modelling needed for sim-to-real transfer.","marker":"[3]"},{"why":"Supplies the base linear velocity estimator trained alongside the policy, required because no motion priors are used.","marker":"[19]"},{"why":"Supplies the reward structure, MDP formulation, and PPO hyperparameter starting point adapted for this work.","marker":"[25]"},{"why":"Provides the RaiSim physics engine used to generate, simulate, and admit the randomised robot set.","marker":"[22]"}],"fun_headline_variants":["One policy zero-shot transfers across 12-50 kg quadrupeds","Dynamics-conditioned policy cuts velocity error 30% on ANYmal C","Reference-free PAL adapts to unseen quadrupeds up to 50 kg","Morphology-aware conditioning beats temporal encoding on ANYmal C"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-picked kinematic and dynamic randomisation ranges, together with the rule that a generated robot must stand collision-free for 2 seconds to be admitted, cover the real dynamics of every target robot, including the ANYmal C used in the hardware tests.","fun_headline_variants_meta":{"raw":{"variants":["One policy zero-shot transfers across 12-50 kg quadrupeds","Dynamics-conditioned policy cuts velocity error 30% on ANYmal C","Reference-free PAL adapts to unseen quadrupeds up to 50 kg","Morphology-aware conditioning beats temporal encoding on ANYmal C"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00153,"raw_usage":{"total_tokens":6140,"prompt_tokens":974,"completion_tokens":5166,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":5087}},"tokens_in":590,"tokens_out":5166,"duration_ms":31602,"temperature":1.0,"reasoning_tokens":5087,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:07:32.031380+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same two architectures on a reference set that leaves out ANYmal C-like body masses and leg lengths, then deploy on ANYmal C and measure heading velocity RMSE. If the error stays near the full-reference values of 0.1652 for PAL and 0.0885 for MorAL instead of degrading toward the single-reference values of 0.2193 and 0.1339, the paper's claim that diverse reference modelling drives the improvement would be wrong; the expected contrast is given in the paper's own Table V.","supporting_citations":[{"cited_title":"MetaLoco: Universal Quadrupedal Locomotion with Meta-Reinforcement Learning and Motion Imitation","cited_arxiv_id":"2407.17502","evidence_quote":"Motivates the GRU-based Dynamics Inference Module used to embed robot dynamics from state history."},{"cited_title":"Learning Low-Frequency Motion Control for Robust and Dynamic Robot Locomotion","cited_arxiv_id":"2209.14887","evidence_quote":"Supplies the reward structure, MDP formulation, and PPO hyperparameter starting point adapted for this work."},{"cited_title":"RaiSim – v1.1.7 documentation,","cited_arxiv_id":null,"evidence_quote":"Provides the RaiSim physics engine used to generate, simulate, and admit the randomised robot set."}],"review_version":1}