{"id":"1c485213-18e5-4bac-92c0-3bf6ff4746c0","arxiv_id":"2606.23249","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"LP-NavOA distills a recurrent local planner into a frozen PPO locomotion backbone to enable obstacle bypassing and goal recovery for humanoids under limited perception, improving simulated on-time arrival from 38-40% to 85-97%.","lead":"The paper presents LP-NavOA, a framework that trains a PPO locomotion policy for humanoids and distills a recurrent planner from A-star teachers to handle obstacle avoidance and goal recovery using only local sensing and proprioception. Smart generalists might read it to see how reinforcement learning and distillation can enable map-free navigation for legged robots in cluttered spaces.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Short-range sensing + body-frame goal direction may be insufficient for reliable post-occlusion goal recovery without map or waypoints","rationale":"The reader's weakest assumption is precisely the load-bearing link between teacher rollouts (global map) and student deployment (limited perception). The reported MuJoCo numbers are consistent with the claim only if that transfer succeeds; the abstract supplies no further evidence that would remove the risk.","tokens_in":1838,"tokens_out":299,"duration_ms":21179,"concrete_test":"In the MuJoCo indoor-layout experiments, replace the recurrent planner with an otherwise identical feed-forward network (same inputs, same training data) and recompute on-time arrival; if the metric falls below 60% while brush/contact metrics rise, the recurrence is required for the claimed post-occlusion recovery.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline performance gain (38-40% to 85-97% teacher-calibrated on-time arrival) rests on the recurrent planner generalizing bypassing and recovery from A-star/waypoint teachers that have global information. At runtime the student receives only proprioception, short-range local range sensing, and instantaneous body-frame goal direction. If the recurrent hidden state cannot maintain an internal representation of the occluded goal location, recovery after bypassing fails and the gain collapses to the backbone-only controller. The abstract provides no explicit verification that the distillation transfers this capability under partial observability.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents LP-NavOA, a limited-perception navigation and obstacle-avoidance framework for humanoid robots. It first trains a raycast-conditioned PPO locomotion backbone using a robot-centered circular heading-speed command and safety filter. With the backbone frozen, A-star and waypoint teachers generate rollouts to distill a recurrent local planner that overwrites only the heading command at deployment. At runtime the system uses only proprioception, short-range local range sensing, and body-frame goal direction with no global map or continuous waypoints. In MuJoCo open-wall and indoor layouts the distilled planner achieves obstacle bypassing and post-avoidance goal recovery, raising teacher-calibrated on-time arrival from 38-40% to 85-97% while reducing brush/contact-heavy progress relative to the backbone-only controller. Ablations highlight the importance of dynamic route shaping, teacher-active data collection, and the circular command interface; a Unitree G1 deployment demonstrates hardware executability.","tokens_in":1960,"tokens_out":636,"duration_ms":19786,"significance":"If the central performance claims hold under the stated limited-perception conditions, the work would be significant for humanoid locomotion because it integrates local navigation and obstacle avoidance without explicit global planning or continuous external commands, while preserving an intact whole-body policy. The distillation approach that freezes the backbone and trains only a recurrent heading overwrite is a clean architectural choice that avoids command-tracking mismatches. The reported gains in on-time arrival and the hardware demonstration on a Unitree G1 provide concrete evidence of practicality for cluttered environments.","major_comments":[{"comment":"Abstract, paragraph on MuJoCo results: the headline claim that the distilled planner raises teacher-calibrated on-time arrival from 38-40% to 85-97% is load-bearing for the paper's contribution, yet the abstract supplies no information on the number of evaluation episodes, statistical significance, variance across runs, or the precise definition of the 'teacher-calibrated' metric and 'brush/contact-heavy progress' metric. Without these details the magnitude and reliability of the reported improvement cannot be assessed.","section":"Abstract"},{"comment":"Abstract, runtime usage and MuJoCo results paragraphs: the central assumption that a recurrent hidden state can maintain an internal representation of an occluded goal location (enabling post-avoidance recovery) when the student receives only short-range local range sensing and instantaneous body-frame goal direction is not supported by any explicit verification. No hidden-state analysis, ablation removing recurrence, or controlled occlusion-duration experiments are described, leaving open the possibility that the observed gains collapse to the backbone-only baseline under the partial-observability conditions actually present at deployment.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract states that ablations demonstrate the importance of dynamic route shaping, teacher-active data collection, and the circular command interface, but does not indicate where these ablation results are presented or quantified (e.g., in a dedicated table or figure).","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments. We address each major comment below, agreeing where the manuscript requires clarification or additional content and indicating the planned revisions.","responses":[{"response":"We agree that the abstract would be strengthened by including these supporting details. In the revised manuscript we will expand the relevant abstract paragraph to report that results are averaged over 100 episodes per layout across 5 random seeds (with standard deviations), and we will add concise parenthetical definitions: 'teacher-calibrated on-time arrival' is the fraction of episodes reaching the goal within a time budget derived from teacher performance, while 'brush/contact-heavy progress' measures the share of forward progress accompanied by high contact forces. These quantities are already defined and tabulated in Section 4; the revision will simply surface them in the abstract.","revision_made":"yes","referee_comment":"[Abstract] Abstract, paragraph on MuJoCo results: the headline claim that the distilled planner raises teacher-calibrated on-time arrival from 38-40% to 85-97% is load-bearing for the paper's contribution, yet the abstract supplies no information on the number of evaluation episodes, statistical significance, variance across runs, or the precise definition of the 'teacher-calibrated' metric and 'brush/contact-heavy progress' metric. Without these details the magnitude and reliability of the reported improvement cannot be assessed."},{"response":"The referee is correct that the current manuscript provides no direct verification (hidden-state probing, recurrence ablation, or occlusion-duration sweeps) of the recurrent planner's ability to retain goal information under occlusion. While the existing ablations already isolate the contribution of the recurrent component, we will add in the revision a new analysis subsection that (i) visualizes hidden-state correlations with goal direction during occlusion intervals, (ii) compares recurrent versus feed-forward planner variants, and (iii) reports recovery success as a function of controlled occlusion length. These additions will directly test the partial-observability assumption.","revision_made":"yes","referee_comment":"[Abstract] Abstract, runtime usage and MuJoCo results paragraphs: the central assumption that a recurrent hidden state can maintain an internal representation of an occluded goal location (enabling post-avoidance recovery) when the student receives only short-range local range sensing and instantaneous body-frame goal direction is not supported by any explicit verification. No hidden-state analysis, ablation removing recurrence, or controlled occlusion-duration experiments are described, leaving open the possibility that the observed gains collapse to the backbone-only baseline under the partial-observability conditions actually present at deployment."}],"tokens_in":1630,"tokens_out":549,"duration_ms":21998,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing here is a distillation setup where a recurrent planner trained on A-star rollouts only tweaks the heading of a frozen PPO locomotion policy, and it lifts on-time arrival in sim from ~40% to 85-97%.\n\nWhat the work actually does is keep the whole-body controller untouched while adding a lightweight recurrent module that runs on proprioception, short-range range data, and body-frame goal direction. The ablations on route shaping, active teacher collection, and the circular command interface give concrete evidence that those pieces matter for both training the 3 m/s backbone and for final efficiency. The Unitree G1 run shows the stack executes without continuous external steering, which is a practical check.\n\nThe soft spots sit in the evaluation and the transfer claim. The reported gains come without error bars, run counts, or statistical tests, so the size of the improvement is hard to gauge. More critically, the teachers have global information during data generation, yet the student at deployment sees only local sensing and instantaneous goal direction. Nothing in the abstract confirms that the recurrent state actually carries the occluded goal location forward; if it does not, the post-bypass recovery collapses and the headline numbers shrink to the backbone baseline. That is the load-bearing assumption and it is not directly checked.\n\nThis is useful for groups that already run a stable humanoid locomotion policy and need a drop-in local navigation layer without building a full map or planner. It is worth sending to a serious referee because the architecture is straightforward to reimplement and the hardware result is there, even though the sim claims need tighter variance reporting and a clearer test of memory under occlusion.","headline":"The paper's main move is distilling a recurrent heading-only planner from A-star teachers onto a frozen raycast PPO backbone, which lifts sim arrival rates but leaves the partial-observability transfer untested.","tokens_in":2450,"tokens_out":419,"would_cite":false,"duration_ms":17156,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Freezing a locomotion backbone and distilling a recurrent heading planner lets humanoids navigate and recover goals after occlusion with only short-range sensing.","keywords":["humanoid navigation","obstacle avoidance","limited perception","local planner distillation","recurrent policy","PPO locomotion","whole-body control","MuJoCo simulation"],"falsifier":"A MuJoCo trial in which an obstacle fully occludes the goal, then clears, and the robot with the distilled planner active fails to regain the correct heading direction and reach the goal within the expected time window.","tokens_in":2744,"feed_emoji":"🤖","tokens_out":777,"duration_ms":22134,"temperature":0.7,"pith_summary":"The paper establishes a two-stage approach where a raycast-conditioned PPO locomotion policy is first trained to accept circular heading and speed commands with a safety filter, then frozen while a recurrent local planner is distilled from A-star and waypoint teachers. This planner overwrites only the heading command at runtime, using proprioception, short-range range sensing, and body-frame goal direction without any global map or waypoint stream. A sympathetic reader would care because the method keeps the whole-body policy intact while adding obstacle bypassing and post-occlusion goal recovery, which raises teacher-calibrated on-time arrival from 38-40% to 85-97% and cuts brush or contact-heavy progress in MuJoCo layouts. Ablations highlight the roles of dynamic route shaping, teacher-active collection, and the circular interface for both efficiency and training a 3 m/s backbone. Hardware tests on a Unitree G1 confirm the commands remain executable without continuous external steering.","feed_headline":"Distilled planner lifts humanoid arrival rate from 38% to 85-97%","feed_subtitle":"By overwriting only heading on a frozen backbone it recovers goals after occlusion using short-range sensing and body-frame direction alone.","key_machinery":"The recurrent local planner distilled from teachers to overwrite only the heading command on a frozen whole-body locomotion backbone.","core_discovery":"The central claim is that distilling a recurrent local planner from A-star and waypoint teachers onto a frozen raycast-conditioned PPO locomotion backbone produces obstacle bypassing and post-avoidance goal recovery; at deployment the planner overwrites only the heading command while the whole-body policy stays untouched, yielding 85-97% on-time arrival and lower contact rates in open-wall and indoor MuJoCo scenes when the system receives only proprioception, short-range local range sensing, and body-frame goal direction.","pith_inferences":["The separation of heading overwrite from the frozen whole-body policy could let the same backbone serve other high-level tasks without retraining locomotion.","Because recovery relies only on local sensing, the method might support navigation in environments where building or maintaining a map is impractical.","If the circular command interface generalizes, other command formats could be swapped in to test whether the distillation step remains stable."],"forward_implications":["On-time arrival rises from 38-40% to 85-97% while brush and contact-heavy progress drops relative to the backbone alone.","The system operates at runtime with no global map, waypoint stream, or external planner.","Dynamic route shaping, teacher-active data collection, and the circular command interface are required for the reported navigation efficiency and for training the 3 m/s backbone.","The resulting commands remain executable on Unitree G1 hardware without continuous joystick steering."],"fun_headline_variants":["Distilled planner yields 85-97 percent humanoid arrival rate","Recurrent planner from A-star teachers on frozen PPO backbone","Humanoid navigation reaches 85-97 percent arrival with limited sensing","Distilled local planner enables post-avoidance goal recovery","LP-NavOA uses body-frame direction for 85-97 percent success"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Short-range local range sensing plus body-frame goal direction is enough for reliable post-occlusion goal recovery without a global map or continuous waypoint stream.","fun_headline_variants_meta":{"raw":{"variants":["Distilled planner yields 85-97 percent humanoid arrival rate","Recurrent planner from A-star teachers on frozen PPO backbone","Humanoid navigation reaches 85-97 percent arrival with limited sensing","Distilled local planner enables post-avoidance goal recovery","LP-NavOA uses body-frame direction for 85-97 percent success"]},"model":"grok-4.3","cost_usd":0.005879,"raw_usage":{"total_tokens":2835,"prompt_tokens":752,"num_sources_used":0,"completion_tokens":86,"cost_in_usd_ticks":58787000,"prompt_tokens_details":{"text_tokens":752,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1997,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":752,"tokens_out":86,"duration_ms":13980,"temperature":1.0,"reasoning_tokens":1997,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T08:21:41.658045+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A MuJoCo trial in which an obstacle fully occludes the goal, then clears, and the robot with the distilled planner active fails to regain the correct heading direction and reach the goal within the expected time window.","supporting_citations":[],"review_version":1}