{"id":"db139770-434f-4420-9695-1dcf000199e9","arxiv_id":"2501.04276","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A legged-robot controller that estimates payload and friction online and uses those estimates to switch between agile and recovery policies achieves lower collision rates and higher speeds than non-adaptive baselines.","lead":"This paper presents BAS, a learning framework that lets a quadruped robot avoid obstacles quickly while staying safe when ground friction, payload, or balance changes. It adds a physical-property estimator and a learned safety check to an existing agile locomotive policy, and reports lower collision rates and higher speeds than prior systems in simulation and on a Unitree Go1 robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The safety guarantee depends on RA value sign correctness under estimator error, but the supporting Lipschitz theory in Appendix A is inapplicable to the DRABE-trained function and unverified for the trained policy, leaving the real-world safety claim unsupported.","rationale":"The reader's weakest assumption identifies the validity of the RA network under estimation error and distribution shift as the load-bearing premise. I agree: the safety guarantee is the core novelty over ABS, and it depends on the RA trigger being correct under imperfect estimates. The paper's theoretical claim (Theorem 1) is not actually about the learned DRABE value function and rests on an unverified contraction condition, so it cannot justify the sign-correctness needed for safety. The real-world experiments, while promising, are too few to probe the error space. The concrete perturbation test would directly quantify the RA network's robustness to estimation errors that the paper acknowledges exist. This concern does not reject the empirical results, but it does mean the paper should be conditional on demonstrating that robustness or releasing a certified safety filter.","tokens_in":13794,"tokens_out":6343,"duration_ms":61654,"concrete_test":"In IsaacGym, evaluate the trained BAS policy with the estimator replaced by its true output plus controlled perturbations: (a) Gaussian noise with per-parameter std equal to the measured estimation loss (Table 2) and (b) systematic biases of +-2 kg mass, +-0.2 friction, +-0.02 m CoM shift. Over 1000 episodes per setting, record the RA sign-flip rate, the rate of unnecessary recoveries, and the collision/reach rates. Additionally, compute the empirical Lipschitz constant of the RA network with respect to e via finite differences over the state distribution and check whether gamma(1+L_fpi)<1 holds for the trained policy; if the empirical constant exceeds the theoretical bound or the condition fails, the safety argument is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central safety claim relies on the learned RA network's sign being correct when conditioned on the estimator's output e_hat (Section 3.2). If the RA trigger fires too early or too late under estimation error, the recovery policy cannot prevent collisions. The paper's only theoretical support is Theorem 1 (Appendix A), but the proof analyzes a discounted min-max value function V(s,e)=min_tau max{...} that is not the DRABE fixed point of Equation (2); no argument connects the learned DRABE solution to the sign condition V<=0 <=> s in RA. Moreover, the Lipschitz bound requires gamma(1+L_fpi)<1, which is nowhere verified for the trained policy and is unlikely to hold for a legged system. Empirically, the estimator has nonzero error even after fine-tuning (Table 2: estimation loss 0.323), and the paper admits real-world friction estimates are inaccurate ('we mainly focus on the relative values and trends'). With only 8 trials per real-world condition and average success 81.25% (Table 4), the evidence does not cover the error distribution. Small perturbations in e_hat could flip the RA sign, causing false recovery (agility loss) or missed recovery (collision); the safety claim is therefore not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents BAS, an extension of the Agile But Safe (ABS) framework for quadruped locomotion. BAS trains an agile obstacle-avoidance policy together with a recovery policy, a proprioception-based estimator of physical parameters (payload mass, CoM shift, friction), and a reach-avoid (RA) value network that governs switching to the recovery policy; both the agile policy and the RA network are conditioned on the estimated parameters. An on-policy fine-tuning stage is introduced to reduce distribution shift in the estimator's input history. The authors report simulation comparisons against ABS, RMA-based, Lagrangian, and ablation baselines, plus real-world tests on a Unitree Go1 under payload, friction, and slope variations, claiming a 19.8% increase in speed and a 2.36 times lower collision rate than ABS.","tokens_in":14049,"tokens_out":12239,"duration_ms":119316,"significance":"If the empirical comparisons are reliable, BAS is a practically useful recipe: it combines an interpretable, policy-invariant physics estimator with a learned RA guard and demonstrates real-world transfer across several physical perturbations. The strengths of the paper are the real-robot experiments, the ablation of the fusion and fine-tuning choices, and the explicit conditioning of both policy and guard on estimated parameters. However, the safety claim is not backed by a formal certificate, and the statistical basis for the headline numbers is thin (simulation results are reported without seeds or error bars, and real-world safety uses only 8 trials per condition); the theoretical appendix does not close this gap.","major_comments":[{"comment":"The RA guard is the load-bearing safety mechanism: the system switches to the recovery policy when V_gamma(s, e_hat) > 0. The theoretical support in Appendix A does not establish that this sign decision is correct under estimation error. Theorem 1 proves, modulo the proof issues below, Lipschitz continuity of a discounted min-over-tau value function with respect to e, but the manuscript does not show that this function coincides with the fixed point of the DRABE operator used for training in Eq. (2), nor does it bound the approximation error of the learned network or the effect of replacing e by e_hat. Lipschitz continuity alone gives no sign margin: with the estimation loss of 0.323 reported in Table 2 and the acknowledged inaccuracy of real-world friction estimates in Section 4.4, a perturbation could flip the trigger. I recommend adding an explicit conservatism margin or an empirical sensitivity analysis (e.g., the distribution of RA values near the threshold against estimator error), or softening the safety language.","section":"3.2 and Appendix A"},{"comment":"The quantitative claims rest on experiments without variance or significance reporting. Tables 1 and 2 give point estimates with no indication of how many seeds or randomizations were used, and the text treats small differences as meaningful. In the real world, Table 4 uses 8 trials per condition for safety and 3 trials for agility, without confidence intervals or statistical tests. Under a binomial model, the Slope condition (BAS 5/8 vs ABS 3/8) is not statistically distinguishable, and the aggregate 81.25% vs 34.38% averages over heterogeneous conditions. The headline '19.8% speed increase' and '2.36x lower collision rate' should be reported with intervals or explicitly as descriptive statistics. I also note that BAS-random in Table 1(c) collapses to 100% collision, which underscores that the safety margin under partial estimator failure needs explicit quantification.","section":"4.1, 4.4, Tables 1-4"},{"comment":"The proof of Theorem 1 needs correction. The displayed inequality after Eq. (3) has the second term comparing xi^{pi,e2}_s(kappa) with itself, which cannot yield the claimed bound; the trajectory difference Delta xi(t) is then bounded by an expression involving an undefined quantity p. In Eq. (5), the factor L_fpi appears in the upper bound without justification: solving the displayed maximization gives a bound proportional to gamma^{t*} log(1+L_fpi) / -log(gamma(1+L_fpi)), with no extra L_fpi. More substantively, the condition gamma(1+L_fpi)<1 is assumed but never verified for the trained policies; for a legged system with a high-gain policy this condition is not automatic. These issues do not necessarily invalidate the empirical RA training, but the 'theoretical insights' claimed as contribution 4 should either be fixed or be reframed as heuristic motivation.","section":"Appendix A, Eq. (5)"}],"minor_comments":[{"comment":"Table 1 reports BAS v_peak as 2.70 m/s while Table 2 reports 2.68 m/s for the same configuration; please reconcile the numbers and state the evaluation settings used in each table.","section":"Tables 1 and 2"},{"comment":"The claimed '19.8% increase in speed' should be defined precisely; the agility times in Table 4 (1.74 s vs 1.41 s) imply roughly 23% speed increase if the distance is fixed, so the metric used for the abstract claim should be stated.","section":"Abstract and Table 4"},{"comment":"Please clarify whether the on-policy fine-tuning is performed in simulation only, since real-world deployment has no ground-truth physics available for supervision; if fine-tuning is simulation-only, state this explicitly.","section":"3.3"},{"comment":"For reproducibility, provide the network architectures, estimator history length, alpha schedule, reward weights, PPO hyperparameters, and the DR parameter sampling procedure used for the reported results.","section":"3.1-3.3 and Appendix"},{"comment":"There are several typos and undefined symbols: 'collsion' appears in the Table 1 caption, 'Funtion' appears in Appendix A, and the proof uses p1 and p2 without defining them; please correct these.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The real-robot experiments are valuable and the main empirical direction is plausible, but the paper overstates the safety guarantee: the theoretical appendix does not support sign correctness of the learned RA value under estimation error, and the statistical evidence is thin. I would not require a formal safety certificate for acceptance, but the sign-robustness question and the absence of variance reporting should be addressed before publication. The paper is within scope for the conference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BAS is a useful extension of the ABS line rather than a leap. What's genuinely new is the explicit physics-parameter estimator (mass, CoM, friction) feeding both the agile policy and the reach-avoid value network, plus an on-policy fine-tuning phase that deals with the history shift when recovery and agile policies alternate. The ablation tables make a convincing case that this integration and the fine-tuning matter: fusion interpolation and joint training improve estimation and safety metrics. The real-world payload and friction-change demos are credible, and the estimator's trends are interpretable.\n\nThe soft spots are real but not disqualifying. The safety claim leans on a learned RA network whose sign under estimator error is unverified. Appendix A's Lipschitz theorem analyzes a discounted min-max value function, not the DRABE fixed point actually trained, and the condition gamma(1+L_fpi)<1 is neither verified nor plausibly satisfied. So the theoretical support is mostly decorative. Also, the statistical evidence is thin: no seeds or error bars in the simulation tables, and the real-world 'collision-free' claim comes from 8 trials per condition with an average success of 81.25% — there were collisions even for BAS. The paper itself acknowledges some failures, which is honest, but the abstract's 'collision-free' overstates what is measured.\n\nThat said, these weaknesses are common in this subfield, and the central empirical claim — that conditioning on explicit physics estimates improves adaptive safety and agility relative to ABS and RMA-based baselines — holds up in the reported experiments. The paper deserves a serious referee. I'd ask for error bars, a statement about seeds, a code/data release, and a revised abstract that says 'reduced collision rate' rather than 'collision-free.' The theory appendix should either be connected to the trained value function or explicitly labeled as a heuristic insight.\n\nFor a reading group: worth a look if you care about safe legged locomotion or learned safety filters, but not a must-read.","headline":"A credible incremental extension of ABS with real-world demos, but the 'collision-free' claim outruns the evidence and the theory appendix doesn't cover the learned safety function.","tokens_in":14601,"tokens_out":2729,"would_cite":true,"duration_ms":26669,"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 quadruped can stay fast and safe when physics change, if both its controller and its safety guard are told the current payload and friction.","keywords":["reinforcement learning","adaptive safe control","legged locomotion","reach-avoid value network","physical parameter estimation","online adaptation","quadruped robot","policy switching"],"falsifier":"Take the deployed BAS stack and artificially bias the estimator's mass output by a fixed 2 kg offset while leaving friction and center-of-mass estimates correct; if the collision rate or the false-trigger rate changes materially compared with the unbiased run, then the safety guarantee does not hold under exactly the kind of estimation error the estimator is supposed to tolerate.","tokens_in":13573,"feed_emoji":"🤖","tokens_out":5055,"duration_ms":45846,"temperature":0.7,"pith_summary":"The paper claims that a legged robot can stay both fast and collision-free when the physics it walks on change, if the control policy and the safety guard both receive live estimates of the physical parameters. To do this, the authors extend the Agile But Safe (ABS) pipeline with an estimator that reads a short history of the robot's own motion and infers payload mass, center-of-mass shift, and ground friction. The agile policy and the reach-avoid value network are conditioned on those estimates, and an on-policy fine-tuning step corrects the estimator for the distribution shift caused by switching between agile and recovery policies. If the claim holds, the robot should navigate slippery floors, added payloads, and slope changes at higher speed and with fewer collisions than non-adaptive or latent-adaptive baselines. The reported real-world result is a 19.8% speed increase and a 2.36 times lower collision rate than ABS.","feed_headline":"Adaptive safety boosts quadruped speed 19.8%, cuts collisions 2.36x","feed_subtitle":"Conditioning control and safety guard on live physics estimates keeps the robot fast and collision-free.","key_machinery":"The load-bearing mechanism is the parameter-conditioned reach-avoid (RA) value network, a learned approximation of the time-discounted reach-avoid value function $V^\\pi_\\gamma(s,e)$ governed by the Discounted Reach-Avoid Bellman Equation (DRABE). The RA network outputs a value whose sign decides whether the agile policy keeps control or hands over to a recovery policy; conditioning it on estimated physics parameters makes the switch adapt to the current situation. The second component is the proprioceptive physics-parameter estimator, trained jointly with the agile policy via a fusion interpolation between ground-truth and estimated parameters, and fine-tuned on-policy to reduce the history distribution shift caused by policy switching. A Lipschitz-continuity argument for $V^\\pi_\\gamma$ with respect to environment parameters $e$ is given to justify that small estimation errors do not cause the safety value to jump discontinuously.","core_discovery":"On the paper's own terms, the central discovery is that explicit, policy-invariant physical parameters, estimated online from proprioception, are a better conditioning signal for a reach-avoid safety guard than the latent dynamics encodings used by adaptive locomotion policies. Conditioning both the agile goal-reaching policy and the learned reach-avoid value function on these estimated parameters makes the safety trigger fire earlier when payload is heavy and later when it is light, matching the intuition that heavier payloads are more dangerous. The paper further claims that the same safety improvement transfers to the real world, where BAS outperforms ABS and an RMA-plus-Lagrangian baseline on both agility and safety metrics across payload, friction, and slope trials.","pith_inferences":["A testable extension would be to feed the same estimated parameters into a model-based controller or a Hamilton-Jacobi reachability solver, not just a learned value network, to see whether the empirical safety improvement comes from the conditioning signal or from the learned value representation itself.","The paper's own BAS-random ablation, where estimated parameters are replaced with random numbers, collapses success to zero; this suggests the framework is deliberately non-conservative, so a deployment that wants robustness to estimator failure would need a separate monitor that detects abnormally large estimation error.","Because the reach-avoid value function is Lipschitz in the estimated parameters, one could in principle compute a worst-case safety margin from the estimator's error bound; the paper does not do this, but the theory presented makes such a bound a natural next step."],"forward_implications":["If BAS is correct, adaptive safe locomotion no longer requires knowing the physics in advance: the robot infers payload, friction, and center-of-mass shift from its own motion history, so deployment to an unseen floor or load can reuse the same trained stack.","Because the safety guard is parameter-conditioned, the robot can trade off agility and caution online; a heavy payload makes the recovery policy intervene earlier, and removing the payload relaxes the guard without retraining.","The on-policy fine-tuning phase implies that estimator accuracy is a bottleneck: after fine-tuning, estimation loss drops from 0.570 to 0.323 and collision rate from 3.10% to 1.11%, suggesting that fidelity of the estimated parameters directly controls safety performance.","In simulation, BAS reaches the goal in 93.84% of dynamic-environment episodes versus 63.83% for ABS, so the claimed improvement is not only about avoiding collisions but about completing the task."],"supporting_citations":[{"why":"Supplies the ABS agile/recovery policy structure and the reach-avoid network training recipe that BAS extends.","marker":"He et al., 2024b"},{"why":"Provides the DRABE contraction-mapping formulation used to learn the reach-avoid value function with convergence guarantees.","marker":"Hsu* et al., 2021"},{"why":"Defines the reach-avoid sets, safe set, and backward reachable set that frame the safety objective.","marker":"Bansal et al., 2017"},{"why":"Inspires the fusion interpolation and joint training of the control policy with the estimator.","marker":"Ji et al., 2022"},{"why":"Supplies RMA, the latent-adaptation baseline that BAS is compared against.","marker":"Kumar et al., 2021"},{"why":"Inspires the action-distillation baseline and the practice of distilling a privileged teacher policy.","marker":"Lee et al., 2020"},{"why":"Provides the earlier Lipschitz-continuity result for the value function that this paper extends to environment parameters.","marker":"Li et al., 2024a"}],"fun_headline_variants":["Quadruped robot learns adaptive safety, speeds up 19.8%","Speed + safety: online physics estimation cuts collisions 2.36x","Adaptive control bridges agility and safety in legged robots","Learn agility without crashing: a physical-parameter-aware safety net","Faster and safer: conditioning on live physics beats latent encodings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned reach-avoid value network, trained in simulation with ground-truth parameters, continues to act as a correct safety trigger when it is fed the estimator's real-world outputs; if estimation error pushes the value across the switching threshold at the wrong moment, the recovery policy cannot prevent the collision.","fun_headline_variants_meta":{"raw":{"variants":["Quadruped robot learns adaptive safety, speeds up 19.8%","Speed + safety: online physics estimation cuts collisions 2.36x","Adaptive control bridges agility and safety in legged robots","Learn agility without crashing: a physical-parameter-aware safety net","Faster and safer: conditioning on live physics beats latent encodings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1297,"prompt_tokens":964,"completion_tokens":333,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":241}},"tokens_in":580,"tokens_out":333,"duration_ms":3335,"temperature":1.0,"reasoning_tokens":241,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:37:03.859336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the deployed BAS stack and artificially bias the estimator's mass output by a fixed 2 kg offset while leaving friction and center-of-mass estimates correct; if the collision rate or the false-trigger rate changes materially compared with the unbiased run, then the safety guarantee does not hold under exactly the kind of estimation error the estimator is supposed to tolerate.","supporting_citations":[{"cited_title":"Hamilton-Jacobi Reachability: A Brief Overview and Recent Advances","cited_arxiv_id":"1709.07523","evidence_quote":"Defines the reach-avoid sets, safe set, and backward reachable set that frame the safety objective."}],"review_version":1}