{"id":"52e60a5f-8f0c-4ddc-9991-b3a432a30e0a","arxiv_id":"2508.06229","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"REBot uses two reinforcement-learned policies in a finite-state machine so a quadruped can dodge fast-moving obstacles and stabilize itself afterward.","lead":"A new control framework helps quadrupedal robots perform fast, reflex-like dodges when obstacles approach in under 1.5 seconds, then recover their balance. The system was tested in simulation and on a real Unitree Go2 robot, though real-world success rates were moderate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Perception is assumed, not validated: real-robot DOA results rely on OptiTrack ground-truth injection, so the claimed reflexive-evasion capability under <1.5 s is not established under realistic sensing.","rationale":"The reader's weakest assumption—perfect state information via motion capture—is exactly the point on which the central claim hinges. The paper is transparent about this limitation, which is a point in its favor, but transparency does not remove the dependency. The FSM's switch into avoidance and the learned policy's actions both consume obstacle position and velocity; if those quantities are wrong or late, the entire reflexive-evasion behavior is compromised. No experiment in the paper varies observation noise or latency, and the real-robot evaluation does not measure actual reaction-time from detection to action, so the headline 'under 1.5 s' is not directly evidenced on hardware. The 56% hardware ASR further shows that the sim-to-real gap is already non-trivial even with oracle perception. Because the reader's conditional verdict already reflects this concern, my stress-test does not move the verdict; it reinforces it and proposes a concrete noise/latency experiment that would settle whether the assumption actually undermines the claimed capability.","tokens_in":13239,"tokens_out":7807,"duration_ms":89748,"concrete_test":"Run the deployed avoidance policy in Isaac Gym (or as a sim2real test) with p_o and v_o corrupted by zero-mean Gaussian noise with sigma in {0.05, 0.1, 0.2} m and observation latency in {0.02, 0.05, 0.1} s, for T_react in 0.5-1.5 s; measure ASR/RSR. If ASR/RSR stay within about 5% of Table I's 0.65/0.59, the perception assumption is benign; if they degrade materially, the claimed real-time evasion is not robust to realistic sensing and the central claim must be re-scoped to a controller with oracle state.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix III states the robot 'directly receives ground truth position and velocity information for both itself and the obstacles at each control step' from OptiTrack; the FSM trigger in Sec. IV-A and the avoidance policy consume p_o and v_o. The conclusion concedes 'we leave precise obstacle position perception as an assumption.' This is load-bearing because the abstract and introduction claim real-time instantaneous DOA on a robot, but no experiment—simulation or hardware—varies observation noise or latency. A policy trained and evaluated on oracle state can fail when perception is noisy or delayed (e.g., 5 cm error, 50 ms latency), and the 56% hardware ASR already indicates Sim2Real fragility. Therefore the central claim is conditional on an unvalidated perception layer, not solely on the control contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces REBot, a finite-state-machine (FSM) control framework for quadrupedal robots that combines a reinforcement-learned avoidance policy and a recovery policy to handle dynamic obstacles under short reaction times (<1.5 s). The avoidance policy is trained in Isaac Gym with a two-stage curriculum (static then moving obstacles) and includes avoidance, regularization, and adaptive reward components; the recovery policy stabilizes the robot after evasive maneuvers. The authors evaluate REBot in simulation against ABS and RRL baselines, report avoidance success rate (ASR), recovery success rate (RSR), maximum joint power (MJP), and avoidance moving distance (AMD), and conduct ablations removing the recovery stage, first curriculum, and adaptive reward. A real-robot demonstration on a Unitree Go2 uses OptiTrack motion capture to feed ground-truth robot and obstacle states directly to the controller; real-world ASR is reported as 56% and RSR as 53%.","tokens_in":13508,"tokens_out":4243,"duration_ms":48400,"significance":"If the claims held fully, REBot would be a useful step toward reactive, reflex-like obstacle avoidance for legged robots, complementing navigation-based replanning methods. The paper's strengths include a clearly formulated problem, a structured FSM integrating two learned policies, ablation studies substantiating the contributions of the recovery stage, curriculum learning, and adaptive reward, and honest acknowledgment of limitations in the conclusion. The empirical evidence is, however, weakened by the reliance on ground-truth state injection from an external motion-capture system in the real-robot experiments, by the absence of statistical error bars or trial counts, and by the underspecification of reward coefficients and FSM thresholds. As the real-world success rates are far below simulation, the 'robust' and 'real-time' system-level claims are not yet established. These issues are load-bearing but appear fixable with additional experiments and reporting, hence the recommendation for major revision.","major_comments":[{"comment":"The real-robot demonstration injects ground-truth position and velocity of both the robot and the obstacle from OptiTrack at every control step. The conclusion states that precise obstacle perception is left as an assumption. Because the FSM trigger in Sec. IV-A and the avoidance policy consume p_o and v_o, the central real-time DOA result is conditional on an oracle perception layer. No experiment varies observation noise or latency, and Table V randomizes obstacle initial conditions rather than measurement noise. Please either add robustness experiments with noisy or delayed obstacle observations or explicitly scope the contribution as control-only in the abstract and introduction.","section":"Appendix III, Sec. VI, Conclusion"},{"comment":"All simulation metrics are reported as single numbers without confidence intervals, numbers of trials, or random seeds. The ablation differences are small (e.g., ASR 0.65 vs. 0.59 for 'w/o adp.' in the 0.5–1.5 s row) and may be within run-to-run variance. The real-robot ASR (56%) and RSR (53%) also lack trial counts or per-scenario breakdowns. Please provide statistics over multiple seeds and trials, and specify how success is judged in hardware experiments.","section":"Tables I and II, Sec. VI"},{"comment":"Reward coefficients for r_avoidance, r_regularization, and r_adaptive are never specified, nor are the hyperparameters λ and η in the threat reward r_threat. The FSM instability thresholds θ_th, qdot_th, and h_th in Sec. IV-A are also absent. This prevents reproduction of the training pipeline and makes the claim of 'carefully designed' rewards unverifiable. Please list all reward weights and threshold values, or release the training code.","section":"Sec. IV-B and Appendix I"},{"comment":"The baselines ABS and RRL are not adapted to the task: ABS targets static obstacles and RRL targets UAV dynamics. A comparison against a legged reactive planner or an RL policy trained on the same observation space without the FSM/recovery stage would better support the claim that the proposed structure, rather than the training setup, yields the improvements. The ablations in Table II partly address this, but the abstract's 'notable improvements' rests on a weaker baseline comparison.","section":"Sec. V-B, Table I"}],"minor_comments":[{"comment":"The transition condition for switching to the avoidance stage is garbled by missing symbols: '�v� , p� � p� � > �' should be the dot product of obstacle velocity and relative position. Please fix the equation rendering.","section":"Sec. IV-A"},{"comment":"Several equations contain placeholder symbols, e.g., rdistance = � ����� �d(p_o, OBB) − r_o� and rcollision = ��c �� � ��c ��. Please replace with standard mathematical notation.","section":"Sec. IV-B"},{"comment":"The metric 'MJD' appears where the text should refer to 'MJP' (maximum joint power) or 'AMD' (avoidance moving distance); the boundary description 'MJD below 300 W' indicates a typo. Also ensure consistent use of MJP/AMD throughout Figures and text.","section":"Sec. V-C"},{"comment":"There is a missing space before 'Under the real-world test conditions' in the last paragraph. Additionally, the sentence 'the REBot system achieves an ASR of 56% and an RSR of 53%' would benefit from specifying the total number of trials and the definition of recovery success on hardware.","section":"Sec. VI"},{"comment":"The red dashed region labeled 'reflexive evasion' should be defined in the caption; the reader must infer the reaction-time intervals from Table I.","section":"Fig. 5 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid empirical study but currently overclaims system-level capability given the Oracle-vision hardware setting and low real-world success rates. The fit with the journal is acceptable, but the authors should be asked to either incorporate onboard perception or clearly delimit the claims, and to add statistical rigor and full training details. If the perception assumption cannot be relaxed, a control-focused framing with a more modest real-world claim would be more appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the FSM control structure: an avoidance policy and a recovery policy trained with a two-stage curriculum and adaptive rewards, targeting sub-1.5s reaction times. That specific integration is new for quadrupedal reflexive evasion, and the simulation ablations give it credible support. Removing the recovery stage drops success rates by about 20% in the reflex regime; removing the adaptive reward collapses gait diversity; the curriculum helps. The authors also report real hardware trials, which is more than many RL locomotion papers bother with, and they state their limitations plainly. That honesty is worth crediting.\n\nThe soft spots are real and roughly as the reader framed them. The baselines are weak: ABS is not built for dynamic obstacles, and RRL is a UAV policy with simplified rigid-body dynamics, so beating them mostly shows the problem is hard, not that REBot is strong. The real-world ASR of 56% and RSR of 53% sit uneasily with the word \"robust\" in the abstract. There are no error bars, no trial counts, no code or data, and several key thresholds and reward coefficients are undisclosed.\n\nThe biggest issue is the perception assumption. Appendix III says the robot receives ground-truth position and velocity for itself and the obstacle at every control step from OptiTrack, and the conclusion concedes that precise obstacle perception is left as an assumption. Since the avoidance policy and the FSM trigger consume p_o and v_o directly, the entire reflexive-evasion claim is conditional on having perfect state. No experiment varies observation noise or latency, and the 56% hardware ASR already hints that the policy is fragile to reality. The stress-test note lands: the paper's central promise is a robot that reflexively avoids in real time, but what is actually validated is a control policy that works given oracle perception. That is a meaningful distinction.\n\nStill, the paper is not incoherent and does not overclaim relative to its own evidence. It is a solid empirical study of a control architecture, with honest limitations and a reproducible-in-principle simulation setup.\n\nMy recommendation: send it to peer review. A good referee will push for code or data, for more hardware trials, for stronger baselines, and for at least one experiment with noisy or delayed perception. If those are addressed, the paper would be a useful reference for legged-robot safety. As it stands, I would not put it in a reading group as a must-read, and I would not cite it in my own work until the perception gap is closed.","headline":"Useful control-system contribution with an honest but load-bearing perception assumption; deserves review, not unconditional acceptance.","tokens_in":13941,"tokens_out":1270,"would_cite":false,"duration_ms":15667,"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 learned reflex, not trajectory replanning, lets a quadruped avoid obstacles with under 1.5 seconds of warning.","keywords":["quadrupedal robots","dynamic obstacle avoidance","reflexive evasion","reinforcement learning","finite-state machine","recovery control","sim-to-real","Unitree Go2"],"falsifier":"Run the same REBot policies on a quadruped using only onboard sensing, removing the motion-capture ground-truth injection, and measure avoidance success at reaction times between 0.5 and 1.5 s; if the success rate falls to the level of the navigation baselines, the claimed reflexive evasion was an artifact of perfect state information.","tokens_in":13167,"feed_emoji":"🤖","tokens_out":10880,"duration_ms":106045,"temperature":0.7,"pith_summary":"The paper proposes that quadrupedal robots can handle obstacles that arrive too fast for path replanning by acting reflexively, and that the reflex can be learned. REBot is a finite-state machine with three stages: normal operation, a learned avoidance policy that dodges the incoming obstacle, and a learned recovery policy that regains balance afterwards. The authors train both policies with a standard policy-gradient method in a GPU simulator, using a two-stage curriculum from suddenly appearing static obstacles to moving ones, plus regularization and adaptive rewards. They report that this split—evade first, recover second—raises avoidance success over navigation-based baselines when reaction time is below 1.5 seconds, and that the same policies transfer to a real quadruped. If the system is right, instantaneous dynamic obstacle avoidance does not require replanning speed; it requires a fast evasive skill and a separate stabilization skill.","feed_headline":"Learned reflex lets a quadruped dodge in under 1.5 seconds","feed_subtitle":"A two-stage reflex dodges then recovers balance, beating replanning when obstacles arrive too fast to navigate.","key_machinery":"The load-bearing object is the three-stage finite-state machine of the REBot system. The Avoidance stage is driven by a policy-gradient-trained actor network that outputs joint target angles; the Recovery stage is a second trained network activated by thresholds on body orientation, joint velocity, and base height. The FSM's work is to decouple the reflex from balance restoration: the avoidance policy can prioritize speed at the expense of stability, while the recovery policy restores normal function. A two-stage curriculum—static obstacle appearing at random position, then moving obstacle on a directed trajectory—and reward terms for trot symmetry, energy, contact smoothness, diversity, and","core_discovery":"The paper's central claim is that reflexive evasion should be a distinct control layer for legged robots: when reaction time is under 1.5 s, the robot should not try to recompute a safe path, it should immediately execute a learned evasive gesture. REBot implements this as a finite-state machine. On detecting an obstacle moving toward the robot, it enters the Avoidance stage; a neural-network policy trained by reinforcement learning maps proprioceptive state plus obstacle position and velocity to joint targets, producing behaviors like jumping backward for frontal threats and crouching for overhead ones. When the dodge leaves the body tilted, joint velocities extreme, or base height low, REB","pith_inferences":["Extension: replacing the motion-capture ground truth with onboard perception is the natural next stress test; REBot's control layer is agnostic to where obstacle state comes from, so the perception module, not the reflex policy, is the remaining bottleneck.","Extension: the observed asymmetry—backward jumps are easier for the Go2's leg morphology—suggests morphology-aware action priors or asymmetric rewards could balance evasion capability across directions for other quadruped designs.","Extension: the reflex-versus-navigation split implies a hybrid controller that uses reaction time or threat level to switch between REBot's evasion policy and a replanning layer, which would extend the finite-state machine to continuously varying threat levels.","Extension: a falsifiable prediction of the design is that failures in a mixed-speed obstacle course should concentrate in perception or actuation, not in the decision layer, because the FSM explicitly covers the short-reaction-time regime."],"forward_implications":["At reaction times between 0.5 and 1.5 s, REBot reports an avoidance success rate of about 0.65, versus 0.11 and 0.09 for the ABS and reactive-RL baselines, so a reflex layer can outperform replanning exactly where replanning is hardest.","Removing the recovery stage substantially reduces post-avoidance stabilization in the reflex regime, so the evasive dodge and the balance recovery should be designed together, not as one black-box policy.","The learned policy produces direction-dependent behaviors—jumping away for frontal and lateral threats, crouching for overhead ones—so one reactive controller can cover a family of evasive maneuvers rather than a single predefined motion primitive.","In the longer reaction-time regime (1.5–4.0 s), REBot still beats the baselines while using less maximum joint power and less base displacement, indicating the reflex architecture does not sacrifice planned, efficient avoidance.","The reported real-robot success rates (ASR 56%, RSR 53%) show the trained policies transfer beyond simulation, making the framework a candidate for safe legged robots in human environments."],"supporting_citations":[{"why":"Agile But Safe; the navigation-based DOA baseline REBot is compared against and outperforms at short reaction times.","marker":"[9]"},{"why":"Reactive RL for UAVs; the end-to-end evasion baseline that anchors the learned-reactive control comparison.","marker":"[31]"},{"why":"Proximal Policy Optimization; the algorithm used to train both the avoidance and recovery policies.","marker":"[15]"},{"why":"Massively parallel GPU reinforcement-learning training environment used for the curricula and ablations.","marker":"[37]"},{"why":"Diversity skill-learning objective that the adaptive reward's diversity term draws on to prevent convergence to a single evasion gait.","marker":"[36]"}],"fun_headline_variants":["Reflexive dodging beats replanning for fast obstacles","Two-stage reflex: dodge then recover for quadruped","REBot uses learned evasive gestures for instant dodging","Learned reflex dodges then recovers balance for quadrupeds"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The robot receives exact ground-truth position and velocity of itself and the obstacles at every control step from a motion-capture system, so the central evasion result assumes such perception is available; the paper leaves precise obstacle perception to future work.","fun_headline_variants_meta":{"raw":{"variants":["Reflexive dodging beats replanning for fast obstacles","Two-stage reflex: dodge then recover for quadruped","REBot uses learned evasive gestures for instant dodging","Learned reflex dodges then recovers balance for quadrupeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000473,"raw_usage":{"total_tokens":2158,"prompt_tokens":690,"completion_tokens":1468,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":434,"completion_tokens_details":{"reasoning_tokens":1410}},"tokens_in":434,"tokens_out":1468,"duration_ms":10938,"temperature":1.0,"reasoning_tokens":1410,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:49:36.434491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same REBot policies on a quadruped using only onboard sensing, removing the motion-capture ground-truth injection, and measure avoidance success at reaction times between 0.5 and 1.5 s; if the success rate falls to the level of the navigation baselines, the claimed reflexive evasion was an artifact of perfect state information.","supporting_citations":[{"cited_title":"Agile but safe: Learning collision-free high-speed legged locomotion,","cited_arxiv_id":null,"evidence_quote":"Agile But Safe; the navigation-based DOA baseline REBot is compared against and outperforms at short reaction times."},{"cited_title":"Flying in highly dynamic environments with end-to-end learning approach,","cited_arxiv_id":null,"evidence_quote":"Reactive RL for UAVs; the end-to-end evasion baseline that anchors the learned-reactive control comparison."},{"cited_title":"Prox- imal policy optimization algorithms,","cited_arxiv_id":null,"evidence_quote":"Proximal Policy Optimization; the algorithm used to train both the avoidance and recovery policies."},{"cited_title":"Learning to walk in minutes using massively parallel deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Massively parallel GPU reinforcement-learning training environment used for the curricula and ablations."},{"cited_title":"Diversity is all you need: Learning skills without a reward function,","cited_arxiv_id":null,"evidence_quote":"Diversity skill-learning objective that the adaptive reward's diversity term draws on to prevent convergence to a single evasion gait."}],"review_version":1}