{"id":"90cdd6e8-8a56-4d2c-a2d4-15c5e05449c0","arxiv_id":"2607.15701","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Reinforcement learning that adjusts obstacle-inflation radii in a visibility-graph planner improves humanoid navigation robustness under delay and noise, beating a static MPC baseline and an end-to-end RL policy in a single simulated scenario and a qualitative hardware test.","lead":"RAVEN is a navigation system for humanoid robots that uses reinforcement learning to tweak how obstacles are inflated in a visibility-graph planner, while a model-predictive controller handles the actual trajectory. It aims to make humanoid navigation more robust to control delays and sensor noise without giving up interpretable safety constraints.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-scenario evaluation with no variance or held-out tests makes the quantitative performance claims unsupported.","rationale":"The reader identified sim-to-real transfer as the weakest assumption, which is a real concern because the hardware section (V-B) provides only qualitative trajectory comparisons with no success rates or error metrics. However, I think the more load-bearing concern is upstream: even the simulation evidence, which is the primary quantitative support for the central claim, is statistically uncharacterized and confined to a single scenario. If the simulation results lack variance reporting or held-out environments, the claimed advantages over baselines may not be reliable, and the transfer question becomes moot. The hardware experiment uses essentially the same obstacle layout as the simulation, so it cannot compensate for the lack of generalization evidence. I therefore partially agree with the reader: both concerns reflect thin evidence, but the single-scenario, no-variance simulation evaluation is the more fundamental gap. The paper's architecture is coherent and the idea is plausible, so conditional acceptance remains appropriate until the proposed held-out evaluation is performed. I would keep the verdict CONDITIONAL rather than rejecting, because the concern is addressable with additional experiments rather than indicating a fundamental flaw.","tokens_in":10990,"tokens_out":8145,"duration_ms":71151,"concrete_test":"Evaluate RAVEN, the fixed-inflation MPC baseline, and the pure RL baseline on 20 held-out layouts (randomized positions of 3 obstacles with radius 1.0 m on the same field), with 10 episodes each, under 0.06 s delay. Report mean ± std for path length, completion time, and max penetration depth, and compute paired bootstrap 95% confidence intervals for RAVEN minus each baseline. If any interval includes zero, the claimed advantages are not statistically robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests entirely on Table V (Section V-A), which reports performance for a single fixed navigation scenario: one start pose, one goal pose, and three obstacles at fixed positions. The table lists 'Average' metrics but provides no number of evaluation episodes, no standard deviation, and no significance test. The only other validation, the hardware experiment (Section V-B), uses nearly the same obstacle arrangement, so it cannot establish generalization to other environments. Because the RL meta-policy is trained on randomized episodes, it may have learned a generalizable inflation strategy; however, the paper offers no held-out evaluation to demonstrate this. The observed margins (e.g., 9.33 m vs 9.80 m path length, 11.58 s vs 12.21 s completion time under 0.06 s delay) could be within run-to-run variance. Without statistical characterization, the conclusion that RAVEN is an effective and robust alternative to end-to-end learning is not supported by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RAVEN is a hierarchical navigation framework for humanoid robots. A PPO meta-policy observes robot, goal, and obstacle states and outputs per-obstacle inflation radii used by the DA VG visibility-graph planner; the resulting shortest path is tracked by a cf-MPC layer with linearized obstacle constraints and a low-level locomotion policy. The paper compares RAVEN with a fixed-radius DA VG-cfMPC baseline and an end-to-end RL baseline under zero delay and 0.06 s delay in a single simulation scenario (Table V), and reports a qualitative hardware demonstration on the T1 robot. The central claim is that RL-adapted geometric graph construction is more robust and more efficient than both static MPC and pure RL while retaining interpretability.","tokens_in":11209,"tokens_out":4568,"duration_ms":40716,"significance":"The idea of using RL to modify obstacle inflation in a visibility graph is a genuinely different integration point from prior work that tunes MPC cost weights or generates subgoals; it has the potential to be interpretable and sample-efficient because the graph shortest-path structure provides a planning prior. Strengths include a JAX/GPU implementation with parallel training, a deliberate attempt to equalize total training time between RAVEN and the pure RL baseline, and a real-hardware deployment with a low-level locomotion policy. No circularity is apparent: the learned inflation behavior is not forced by the reward construction. However, the evidence base is too narrow to support the broad robustness and sim-to-real claims made in the paper.","major_comments":[{"comment":"The quantitative support for the main claim rests on a single scenario: one start pose, one goal pose, and three fixed obstacles. The table reports 'Average' metrics with no number of episodes, no seeds, no standard deviations, and no significance tests. The observed margins (e.g., 9.33 m vs 9.80 m path length and 11.58 s vs 12.21 s completion time under 0.06 s delay) could easily be within run-to-run variance. Since the RL meta-policy is trained on randomized episodes, a held-out, multi-seed evaluation over varied obstacle layouts and start/goal configurations is necessary before claiming superiority.","section":"§V-A, Table V"},{"comment":"The term 'collision-free MPC' is contradicted by the reported 'Ave. Max Obstacle Penetration Depth' of 0.03 m for RAVEN and 0.128 m for the baseline. The constraint in §III-A contains a slack variable δ_j ≥ 0, so it is a soft constraint that permits penetration. The paper should either rename this layer or explicitly frame the guarantee as bounded penetration under a chosen safety margin; otherwise the title, abstract, and conclusion overstate the safety property.","section":"§III-A and §V-A, Table V(b)"},{"comment":"The hardware demonstration is a single qualitative trajectory comparison: no success criteria, no repeated trials, and no quantitative metrics such as completion time, penetration depth, or failure count. The text concludes 'more reliable navigation' and 'consistent performance across simulation and real-world' from Fig. 5, but the figure only shows curves. Additionally, the hardware start pose (3.4, -1.0, 1.57) differs from the simulation scenario in Table V (3.0, -1.0, -1.57), making it unclear what the simulated rollout in Fig. 5 was compared against. This section does not support the sim-to-real transfer claim.","section":"§V-B"},{"comment":"The only classical baseline uses a fixed inflation radius of 1 m, with no sensitivity analysis or systematic tuning. A central claim is that automatic adaptation is preferable to manual tuning, but a more favorable fixed radius, or a small set of tuned radii, might close much of the gap. Reporting an ablation over fixed inflation values, and ideally a comparison against RL-tuned MPC cost weights, would make the contribution of geometric adaptation load-bearing rather than an artifact of a single baseline setting.","section":"§IV-A, §V-A"}],"minor_comments":[{"comment":"The obstacle-avoidance inequality should be numbered, and all symbols (V_{k,j}, R_obs,j, δ_j) should be defined immediately after the equation.","section":"§III-A"},{"comment":"The table should state the number of evaluation episodes and seeds for each method, and the authors should consider reporting standard deviations or confidence intervals.","section":"Table V"},{"comment":"Reference [13] lacks publication venue, year, and bibliographic details; please complete it.","section":"References"},{"comment":"The name is written inconsistently as 'RAVEN' in the title and 'RA VEN' in the body; pick one convention.","section":"Throughout"},{"comment":"The term 'penetration depth' is used without definition; define it consistently with the soft-constraint slack variables.","section":"§V-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's main limitation is not circularity but statistical under-support: the quantitative evaluation is a single scenario with no variance, and the hardware section is qualitative. The 'collision-free' terminology should also be corrected to 'bounded penetration' given the soft MPC constraints. With a strengthened evaluation (multi-seed, varied environments, and quantitative hardware metrics) the contribution could be suitable for a robotics venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea here is real: instead of tuning MPC cost weights or going end-to-end, RAVEN uses an RL meta-policy to adjust obstacle inflation radii in a visibility-graph planner, and a cf-MPC layer tracks the resulting path. That is a genuine departure from the closest prior work (RL-tuned or differentiable MPC, subgoal prediction), and the paper positions it well in the related work. The architecture is coherent, the reward design is sensible, and the training details are concrete enough to reproduce (PPO, Brax, MJX, hyperparameters given). The self-citation of [4] for the base pipeline is legitimate reuse, not padding.\n\nWhat the paper does not yet support is the strength of its own claims. The central evidence is Table V: one fixed start, one goal, three fixed obstacles, no number of episodes, no standard deviations, no significance test. The margins over the RL baseline are small (9.33 vs 9.80 m path length; 11.58 vs 12.21 s under delay), and without run-to-run variance you cannot tell if they are meaningful. The hardware section is a qualitative trajectory comparison using nearly the same obstacle arrangement, so it cannot establish generalization either. The paper also calls the pipeline “collision-free” while reporting a 0.03 m penetration depth; that is an overstatement, even if the depth is small. And there is no baseline against RL-tuned MPC weights, which is the most natural comparator for the claimed advantage.\n\nNone of this kills the core idea. The stress-test note is right that the empirical base is thin, but that is a fixable weakness, not a load-bearing flaw. The approach is interpretable, the learned adaptation is not forced by construction, and the framework is a reasonable answer to a real problem in humanoid navigation.\n\nWho is this for? Robotics researchers working on legged navigation or hierarchical RL-MPC. They would get a clear, if preliminary, demonstration of one way to inject learning into the geometric layer of a classical planner. It deserves a serious referee, not a desk reject—but the referee should insist on statistical evaluation across randomized environments, a quantitative hardware trial, a cost-weight-tuning baseline, and a less absolute safety label. If the authors can supply that, the paper becomes a solid contribution.","headline":"A sensible hybrid-planner idea—RL adapting visibility-graph inflation radii—is let down by single-scenario, no-variance evaluation; still worth a serious referee.","tokens_in":11711,"tokens_out":1490,"would_cite":true,"duration_ms":16390,"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 learning layer that adjusts obstacle inflation in a visibility-graph planner makes a humanoid robot navigate more robustly under control delay and sensor noise than either a fixed-geometry MPC or a pure end-to-end RL policy.","keywords":["humanoid navigation","reinforcement learning","model predictive control","visibility graph","obstacle inflation","sim-to-real transfer","collision avoidance","hierarchical planning"],"falsifier":"Run the trained RAVEN policy on the physical humanoid for repeated trials with the same start, goal, and obstacle layout under the nominal 0.06 s delay, and measure per-trial maximum obstacle penetration and success rate; if the median penetration exceeds the 0.03 m safety bound or trials fail at a rate not observed in simulation, the central robustness claim is falsified.","tokens_in":10875,"feed_emoji":"🤖","tokens_out":5622,"duration_ms":45644,"temperature":0.7,"pith_summary":"The paper proposes RAVEN, a hierarchical navigation system for humanoid robots in which a reinforcement-learning meta-policy does not generate commands but instead chooses per-obstacle inflation radii used to build a visibility graph. By enlarging or shrinking obstacles, the learned policy reshapes the free-space geometry and thus changes which shortest path the graph planner selects, compensating for the overshoot and tracking errors introduced by control delay and estimation noise. A collision-free model predictive controller then tracks the chosen path while explicitly enforcing velocity, acceleration, and obstacle-avoidance constraints. In simulation with a 0.06 s actuation delay, RAVEN keeps maximum obstacle penetration at 0.03 m while achieving the shortest average path (9.33 m) and fastest completion (11.58 s) of the three compared methods, and hardware trials on a bipedal robot show trajectories that transfer more consistently than the pure RL policy.","feed_headline":"RL-adapted obstacle sizes beat MPC and end-to-end RL","feed_subtitle":"Reshaping the planner's free space, not tuning control weights, keeps a biped on track under delay and noise.","key_machinery":"The load-bearing object is the adaptive obstacle-inflation radius in the Dynamic Augmented Visibility Graph. A visibility graph is a graph whose vertices include obstacle vertices and the start and goal, with edges between any two mutually visible vertices; shortest paths through it are piecewise straight and tangent to inflated obstacles. RAVEN lets a reinforcement-learning meta-policy choose each obstacle's inflation radius continuously, so the learned layer effectively morphs the graph's geometry and therefore the topology of the planned path. The collision-free MPC layer tracks the planned trajectory through a quadratic program that enforces velocity and acceleration limits and linearize","core_discovery":"The central claim is that learned adaptation of a geometric planner parameter—obstacle inflation—is an effective, interpretable middle ground between rigid classical planning and monolithic end-to-end reinforcement learning. The meta-policy observes the delayed, noisy robot state and outputs an inflation radius for each obstacle in a visibility-graph planner; larger radii push the shortest path further from obstacles, smaller radii allow it to cut closer. Since the graph always returns the shortest path under the current geometry, the RL search is constrained to a family of shortest-path solutions, which the paper argues both guides exploration and preserves global efficiency. The downstream","pith_inferences":["An implication the paper leaves implicit is that the same mechanism of learning geometric parameters rather than control weights could be applied to other graph-based or sampling-based planners, for example adapting connection radii or sample distributions in a probabilistic roadmap, with similar interpretability.","A natural testable extension is to train the meta-policy with moving obstacles; the learned inflation schedule could then be examined to see whether the robot pre-inflates obstacles on the side where control delay is largest.","The paper's hardware evidence is a single qualitative trajectory comparison; a quantitative multi-trial deployment measuring per-trial penetration depth and success rate would directly substantiate the sim-to-real claim.","Since control delay is the dominant disturbance, a fixed analytic rule that inflates obstacles as a function of current speed and heading rate might capture much of the benefit; comparing RAVEN against such a hand-designed schedule would isolate what the learned adaptation adds beyond a well-chosen static inflation policy."],"forward_implications":["Under 0.06 s actuation delay, RAVEN keeps maximum obstacle penetration at 0.03 m while the fixed-inflation MPC baseline reaches 0.128 m, supporting the claim that geometric adaptation suppresses delay-induced overshoot.","RAVEN's average path length (9.33 m) and time to completion (11.58 s) are better than both the classical baseline (11.25 m, 12.32 s) and the end-to-end RL baseline (9.80 m, 12.21 s), showing that learning over shortest-path structures preserves efficiency while adding robustness.","Because the MPC layer, not a learned network, enforces velocity and obstacle constraints, the framework retains explicit safety guarantees and runs at about 100 Hz, fast enough for real-time humanoid navigation.","The asymmetric actor-critic design—the actor sees delayed noisy states while the critic sees privileged clean states—is intended to make the learned inflation policy transfer from simulation to the physical robot.","The approach suggests that learning at the level of geometric planning parameters, rather than control weights or raw commands, can be a scalable way to add adaptability to classical navigation stacks."],"fun_headline_variants":["RL inflates obstacles to keep humanoids on track","For bipeds, RL retunes the map, not the controller","Learning obstacle sizing beats end-to-end robot policies","RL adapts geometry for delay-robust humanoid navigation","Smarter obstacle inflation improves MPC-guided bipeds"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The learned inflation policy transfers from simulation to the physical robot because the simulation faithfully reproduces the robot's control delay, observation noise, and low-level tracking behavior; if the real tracking-error distribution differs from the simulated one, the claimed real-world robustness is not established.","fun_headline_variants_meta":{"raw":{"variants":["RL inflates obstacles to keep humanoids on track","For bipeds, RL retunes the map, not the controller","Learning obstacle sizing beats end-to-end robot policies","RL adapts geometry for delay-robust humanoid navigation","Smarter obstacle inflation improves MPC-guided bipeds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":1086,"prompt_tokens":806,"completion_tokens":280,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":201}},"tokens_in":550,"tokens_out":280,"duration_ms":3402,"temperature":1.0,"reasoning_tokens":201,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:31:46.269439+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained RAVEN policy on the physical humanoid for repeated trials with the same start, goal, and obstacle layout under the nominal 0.06 s delay, and measure per-trial maximum obstacle penetration and success rate; if the median penetration exceeds the 0.03 m safety bound or trials fail at a rate not observed in simulation, the central robustness claim is falsified.","supporting_citations":[],"review_version":1}