{"id":"dc7a58ac-dbbd-47e4-89b6-5fd28543b7e8","arxiv_id":"2601.15995","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An end-to-end RL quadruped parkour controller that estimates an egocentric polar foothold prior (distance and heading) from depth and proprioception beats Cartesian foothold priors and prior baselines in simulation and real robot trials.","lead":"PUMA is an end-to-end reinforcement-learning controller that lets a quadruped robot use depth images and its own body sensors to estimate simple 'foothold hints' — distance and heading to the next landing spot — and use them to jump across gaps, walls, and stepping stones. If the results hold up, it shows that a compact polar foothold cue can replace heavier hierarchical foothold planning in agile legged locomotion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Foothold-prior recoverability from depth is the load-bearing assumption; oracle-to-predicted gap is untested.","rationale":"The paper is a competent empirical study: it reports per-method trials, includes multiple ablations, and openly documents failure cases in Fig. 7. I considered other candidate concerns—thin real-world statistics (10 trials), unavailable code/hyperparameters, and the absence of an explicit supervised loss for the estimator—but these affect reproducibility/confidence rather than the internal logic of the central claim. The most load-bearing condition is that the egocentric polar prior can be recovered from depth well enough to guide the same behaviors that the oracle prior enables in training. The reader's weakest_assumption identifies this. The design itself acknowledges the risk by annealing from ground-truth to predicted inputs (PAS), and Table III shows the polar prior is easier to regress than Cartesian ones, but no experiment isolates task-level sensitivity to prior error. Fig. 7d is direct internal evidence that the assumption is sometimes violated. My recommended verdict is unchanged from CONDITIONAL: the claim is plausible but not fully secured until the oracle-to-estimator gap is quantified. The proposed test would settle whether this concern is decisive or minor.","tokens_in":10457,"tokens_out":6145,"duration_ms":62086,"concrete_test":"Using the trained PUMA policy in Isaac Gym with identical seeds, run the 80° Surmounting and 80° Wall-assisted Gap terrains under two actor inputs: (A) ground-truth f_t for the entire episode, and (B) the estimator's \\hat f_t. Compute SR/TR over 1,000 episodes. Then, in mode (B), add zero-mean Gaussian noise with std = 0.1×, 0.5×, and 1.0× the reported foothold MSE to \\hat f_t and remeasure. If oracle-mode SR exceeds estimator-mode SR by >5 percentage points, or if 1×-MSE noise drops SR by >10 points, the depth-recovery assumption is confirmed load-bearing; if not, the concern is largely answered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PUMA's central claim—polar foothold priors beat Cartesian priors and PIE/Extreme Parkour—rests on the prior being both informative and deployable. Section III-A constructs ground-truth f_t from privileged terrain knowledge: candidates sampled at 1 m intervals along the commanded direction, filtered by d_edge > d_safe, and explicitly including 'the centers of exploitable inclined walls' (Eq. 1). This is a world-frame oracle label: it requires knowing which surfaces are exploitable and where their centers are. The deployment estimator (Sec. III-B.2) must recover the same four scalars from a 10 Hz depth stream. The paper reports MSE (Table III) but not how task success depends on prior error, and its real-world failure Fig. 7d is literally 'Foothold estimation error causing missed wall contact.' If the learned estimator is the bottleneck, then the apparent advantage of the polar representation may reflect easier regression on the training distribution (Table III) rather than a robust inductive bias at deployment. The training pipeline also gives the critic ground-truth f_t and anneals actor input from f_t to \\hat f_t (Eqs. 2–3), so the policy can latch onto oracle information early; the final policy's dependence on \\hat f_t accuracy is never quantified. This is the most load-bearing soft spot because the headline claim is about real-world agile locomotion, not just regression accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents PUMA, an end-to-end reinforcement-learning framework for quadruped parkour. A depth-plus-proprioception estimator predicts an egocentric polar foothold prior—two forefoot distances and two heading errors—that is annealed from ground-truth to predicted values during training via Probability Annealing Selection (PAS). The policy is trained with PPO and a multi-critic reward structure on three discrete terrain types: stepping stones, wall-assisted gaps, and surmounting platforms. Experiments in Isaac Gym (1,000 trials per configuration) and on a DeepRobotics Lite3 robot (10 trials per method) compare PUMA with Cartesian-prior variants and the PIE and Extreme Parkour baselines. The paper reports higher simulation success rates (94.7–98.7%) and better real-world success for PUMA, attributing the gain to the polar representation of foothold priors.","tokens_in":10861,"tokens_out":6064,"duration_ms":64253,"significance":"If the findings hold, the polar foothold prior is a simple and effective inductive bias for agile legged locomotion, avoiding explicit foothold tracking and hierarchical planning while improving perception-to-control transfer. The simulation protocol is a strength: 1,000 trials per configuration, multiple ablations, and a shared terrain curriculum. The zero-shot real-world deployment across three dynamic tasks is valuable, and the paper honestly lists typical failure cases. However, the real-world evidence is limited to 10 trials per condition, and the gap between oracle-generated foothold priors and depth-only estimation is not directly quantified. The central claim is therefore plausible but not yet fully established.","major_comments":[{"comment":"The real-world evaluation uses only 10 trials per method and reports no confidence intervals or statistical tests. For the decisive comparisons—e.g., PUMA 1.0 vs explicit Cartesian 0.6/0.2 on wall-assisted gap, and 0.8 vs 0.1/0.1 on surmounting—the binomial 95% CI for 10/10 is about [0.69,1.0] and for 8/10 is about [0.44,0.97]. The observed differences are not statistically significant at n=10. The abstract's real-world robustness claim requires more trials, repeated runs, or at least error bars.","section":"Section IV-C, Fig. 5"},{"comment":"The ground-truth foothold prior is generated from privileged terrain knowledge: candidates sampled at 1 m intervals including 'centers of exploitable inclined walls', filtered by d_safe. At deployment, a depth-only estimator must recover the same four scalars from 10 Hz depth images. Table III reports regression MSE but does not quantify how prior error affects task success. Fig. 7d explicitly attributes a failure to 'Foothold estimation error causing missed wall contact'. Without a controlled-noise experiment on f_input_t or an analysis of success as a function of estimator error, the oracle-to-estimator gap is a load-bearing unquantified assumption.","section":"Section III-A (Eq. 1) and III-B.2, Fig. 7d"},{"comment":"The MSE comparison between polar and Cartesian priors is not apples-to-apples. The polar target contains distances (meters) and angles (radians); the Cartesian target contains x,y,z (meters). The scales and units differ, and the reported 'MSE(%)' normalization is not defined. The claim that the polar representation 'achieves significantly higher accuracy' is therefore not supported by the data as presented. To make the representation argument, the authors should report a normalized per-component error on a common scale or compare task success under matched prediction error.","section":"Section IV-B.1, Table III"},{"comment":"The comparison between PUMA and the Cartesian-prior ablations is confounded by reward alignment. The foothold rewards (Eqs. 4–6) are defined on d_L, d_R, and ψ—exactly the components of f_t that are fed to the actor in PUMA. The Cartesian-prior variants receive raw x,y,z coordinates and must internally compute these reward-relevant quantities. The higher success of PUMA may reflect direct reward-feature alignment rather than an intrinsically better prior representation. A control that provides Cartesian variants with polar features computed from their predicted coordinates, or that gives PUMA a Cartesian reward, would isolate the representation effect.","section":"Section III-A/C, Table II"}],"minor_comments":[{"comment":"Clarify how f_input_t, the estimated velocity, and the environment latent are concatenated and fed to the actor. The sentence 'These representations are then concatenated with the current observations' is ambiguous about which inputs go to which network head.","section":"Section III-B.2"},{"comment":"Define T and state its value. In Fig. 6b, define what '5K' and '8K' mean and whether curves are averaged over seeds; add error bars or shading.","section":"Eq. (3), Fig. 6b"},{"comment":"The group weights 'Tracking Group 3', 'Foothold Group 1.5', and 'Regularization Group 1.0' are not explicitly tied to w_task, w_foothold, and w_style in Eq. (7). Specify how group weights and per-term weights combine.","section":"Table I, Eq. (7)"},{"comment":"State unambiguously whether the '10 trials' are per terrain condition or per method overall, and report trial-level outcomes or a table with per-trial results. The captions in Fig. 5 are abbreviated and hard to interpret.","section":"Section IV-C"},{"comment":"No code or exact hyperparameters are released. Provide learning rate, PPO clip, annealing total steps T, d_safe, sparse-reward threshold ϵ, and domain-randomization ranges for reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"This is a well-executed empirical study with a strong simulation protocol, but the real-world evidence and the oracle-to-estimator gap need to be strengthened before the headline claim is fully supported. The confound between the polar representation and the reward design should also be addressed. I see no grounds for rejection, but the revision should include additional real-world trials or statistical analysis, a controlled-noise study on the foothold prior, and a clearer comparison of representation accuracy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: PUMA makes a credible empirical case that a polar foothold prior—two distances plus two headings—is a better representation for single-stage quadruped parkour than Cartesian foothold coordinates. The simulation ablations are solid: 1000 trials per condition, consistent margins over explicit and implicit Cartesian priors, and a fair comparison against PIE and Extreme Parkour using their own reward functions and the teacher policy for Extreme Parkour. The real robot demos add weight, and the failure-case breakdown is honestly done—they even show a case where foothold estimation error causes a missed wall contact.\n\nWhat's genuinely new is the egocentric polar decomposition, combined with PAS annealing and multi-critic training to make the single-stage pipeline converge. The Table III regression MSE data support the claim that the polar form is easier to recover from depth than Cartesian coordinates, and the downstream success rates line up with that. The authors also show that the biggest benefit appears on terrain that requires actively exploiting inclined walls, which is the interesting case.\n\nThe soft spots are real but not fatal. The ground-truth foothold prior is defined using privileged terrain knowledge: candidates sampled at 1 m intervals, filtered by d_safe, and explicitly including the centers of inclined walls. At deployment the estimator has to recover those same targets from a 10 Hz depth stream. The paper reports regression MSE but never shows how much task success degrades as prior error increases. The real-world failure in Fig. 7d is exactly this failure mode, so the concern is not hypothetical. That said, the same recoverability issue applies to the Cartesian baselines, so it doesn't undermine the core comparison—it does temper the real-world robustness claims and the 'exceptional' language in the abstract.\n\nSecond, the real-world evaluation is thin: 10 trials per method, no confidence intervals, and the hardest course has a 0.8 success rate. That's enough for a proof-of-concept but not for strong claims. Third, no code or exact hyperparameters are released, and while the reward table and many settings are listed, independent reproduction would be painful. The PAS finding—that it mainly speeds training rather than improving final performance—is honest and not oversold. The multi-critic ablation shows a clear drop on hard terrain, and the qualitative analysis in Fig. 4 is reasonable.\n\nWho is this for: researchers working on legged locomotion or RL with learned priors. It's a useful representation to build on. I'd want code or a detailed training recipe before citing it heavily, but the direction is sound.\n\nRecommendation: send it to review. It's a competent, well-ablated empirical paper with a clear representation hypothesis. The referee should push for code/data release, more real-world trials with uncertainty quantification, and an explicit analysis of sensitivity to prior error. With those changes, it would be a solid contribution.","headline":"Solid empirical case for polar foothold priors in quadruped parkour, but the oracle-to-depth recoverability gap and thin real-world statistics keep it conditional rather than accept.","tokens_in":11283,"tokens_out":2804,"would_cite":true,"duration_ms":27506,"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":"PUMA shows that an egocentric polar foothold prior — two front-foot distances plus two heading errors — beats coordinate-based foothold targets for end-to-end quadruped parkour, reaching 94.7–98.7% simulation success and near-perfect real-w","keywords":["foothold prior","quadruped parkour","reinforcement learning","egocentric polar representation","visual perception","sim-to-real transfer","single-stage policy training"],"falsifier":"Compare polar and Cartesian priors under oracle execution: feed the ground-truth prior (no estimation error) to the actor for both representations on the wall-assisted gap task. If success rates become equal, the polar format's benefit is purely in easing the regression problem; if polar remains superior, the representation changes what the policy itself can learn. Either result would settle the paper's explanation for its advantage.","tokens_in":10367,"feed_emoji":"🦿","tokens_out":8934,"duration_ms":92256,"temperature":0.7,"pith_summary":"PUMA is an end-to-end reinforcement-learning framework for quadruped parkour over discrete terrain — stepping stones, gaps, and high platforms. Its central move is to represent the next foothold not as an (x, y, z) coordinate to track but as an egocentric polar prior: two front-foot distances to the expected contact point plus two heading errors. This four-scalar cue is predicted from an onboard depth camera and proprioception, then fed to a velocity-tracking actor, guiding posture adaptation (twisting to kick off a wall) without strict foothold tracking. In simulation it reaches 94.7–98.7% success across five terrain configurations, beating Cartesian foothold priors and recent baselines; zero-shot real-world trials succeed in 8–10 out of 10 runs. The paper argues that the format of the guidance signal is a decisive inductive bias for agile legged locomotion.","feed_headline":"Four scalar foothold cues beat coordinates in quadruped parkour","feed_subtitle":"Encoding front-foot targets as polar distance and heading lifts zero-shot parkour to perfect runs on most courses.","key_machinery":"The load-bearing object is the egocentric polar foothold prior, f_t = {d_t^(L), d_t^(R), ψ_t, ψ_{t+1}} — four scalars giving the distances from the left and right forefeet to the expected foothold and the heading errors toward the current and next foothold. Its work is to convert a potentially error-prone 3D foot-coordinate target into a minimal directional-and-distance cue that can guide posture adaptation without dictating exact foot placement, preserving the policy's exploratory freedom. Two mechanisms make the single-stage training work: probability annealing selection (PAS), a cosine schedule that gradually hands control from the ground-truth prior to the predicted prior; and a multi-cr","core_discovery":"The paper's central claim is that egocentric polar foothold priors are a better motion-guide for agile quadruped locomotion than coordinate-based foothold targets. The prior decomposes a foothold target into four scalars: the distances from the left and right forefeet to the expected contact point, and the heading errors toward the current and next foothold. A shared depth-plus-proprioception estimator regresses this prior alongside base velocity and a terrain latent, and the actor uses it to adjust body posture rather than to track an exact foot position. On five discrete-terrain tasks, the polar prior yields simulation success rates of 94.7–98.7%, exceeding explicit and implicit Cartesian","pith_inferences":["The advantage of the polar prior may stem from the information bottleneck it imposes; a direct test would feed ground-truth priors (polar vs Cartesian) at execution time and see if the policy gap persists once estimation error is removed.","The oracle-sampled foothold set — hand-chosen centers of inclined walls filtered by edge distance — could be replaced by a learned affordance estimator, removing the need for terrain-specific annotation and making the prior adaptive to unseen structures.","Because the paper's real-world evaluation uses only 10 trials per condition, a larger deployment study is a natural next test to confirm the robustness claims beyond single-run anecdotes.","The polar representation might generalize to bipedal or industrial robots that must aim feet at discrete features, where the same decomposition of distance and heading could reduce the perceptual burden."],"forward_implications":["End-to-end parkour training is feasible without hierarchical foothold planning or teacher–student distillation: the robot discovers galloping gaits and wall-kick propulsion from a four-scalar cue.","Compact polar guidance roughly halves regression error compared with Cartesian foot coordinates, making sim-to-real transfer more robust with noisy onboard depth.","PAS provides a general recipe for annealing from privileged ground truth to predicted estimates in single-stage RL, accelerating convergence without hurting final performance.","Multi-critic reward separation is necessary on inclined terrain, allowing the policy to sacrifice velocity tracking in favor of momentum-building contact phases."],"fun_headline_variants":["Quadruped parkour: polar cues beat coordinate targets","Four scalar foothold cues trump explicit coordinates in parkour","Polar distance and heading priors drive zero-shot parkour gains","Perception-driven polar priors replace hand-coded footholds in parkour","Polar foothold priors, not coordinates, drive parkour success"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the assumption that an oracle can pre-select the geometrically useful contact points — the centers of exploitable inclined walls, filtered by distance to terrain edges — and that the depth-only estimator can recover those same targets from pixels at deployment; the paper's own failure analysis shows this assumption is sometimes violated.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped parkour: polar cues beat coordinate targets","Four scalar foothold cues trump explicit coordinates in parkour","Polar distance and heading priors drive zero-shot parkour gains","Perception-driven polar priors replace hand-coded footholds in parkour","Polar foothold priors, not coordinates, drive parkour success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001082,"raw_usage":{"total_tokens":4330,"prompt_tokens":677,"completion_tokens":3653,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":3563}},"tokens_in":421,"tokens_out":3653,"duration_ms":28659,"temperature":1.0,"reasoning_tokens":3563,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T08:42:57.331851+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare polar and Cartesian priors under oracle execution: feed the ground-truth prior (no estimation error) to the actor for both representations on the wall-assisted gap task. If success rates become equal, the polar format's benefit is purely in easing the regression problem; if polar remains superior, the representation changes what the policy itself can learn. Either result would settle the paper's explanation for its advantage.","supporting_citations":[],"review_version":1}