{"id":"94e4f4b8-ad73-4286-b7c6-713708301464","arxiv_id":"2603.12842","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A reinforcement-learning controller with sequential goal rewards, future-goal lookahead, and an automatic curriculum lets quadrupeds turn smoothly while running fast and deploys on real robots.","lead":"SmoothTurn trains quadruped robots to change direction smoothly at high speed by looking ahead at future goals instead of treating each target in isolation. If it works as claimed, it would make fast, multi-waypoint navigation more practical for rescue and inspection robots.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Wrong manuscript body supplied; SmoothTurn empirical claims cannot be audited at all.","rationale":"The Reader already correctly diagnosed the manuscript mismatch, set confidence LOW, and left the verdict UNVERDICTED pending the real paper. No additional technical soft spot inside SmoothTurn can be identified because the body is absent; the load-bearing failure is precisely the absence of any verifiable evidence for the empirical claims. Hence the verdict stays UNVERDICTED and agreement is full.","tokens_in":28701,"tokens_out":389,"duration_ms":11704,"concrete_test":"Fetch the actual PDF/source of arXiv:2603.12842. Confirm presence of (i) quantitative tables/figures comparing success rate, path length/efficiency, and turning smoothness vs. single-goal baselines at high speed, (ii) ablations that isolate each of the three proposed components, and (iii) real-robot deployment metrics or videos with onboard sensing. If any of these are missing or fail to support the abstract, the central claim collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim (SmoothTurn yields smooth high-speed sequential turning with emergent momentum control, advance facing, efficient paths, and direct real-robot transfer) rests entirely on the abstract's assertion of 'simulation and real-world empirical results.' The CACHEABLE PAPER SOURCE CONTEXT instead contains the full text of an unrelated Floquet-Keldysh FRG paper (arXiv:2603.12844). Consequently there are no methods details, no baselines, no ablations of the sequential reward / lookahead / curriculum, no quantitative metrics, and no real-robot protocol against which to test whether the weakest assumption (that those three ingredients suffice to overcome single-goal policies) actually holds. The claim is therefore uncheckable.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The submission is presented as SmoothTurn, a learning-based control framework for agile sequential local navigation on quadrupedal robots. The abstract claims that a sequential goal-reaching reward, a future-goal lookahead observation, and an automatic goal curriculum yield smooth high-speed turning with emergent momentum control, advance facing, and efficient paths, with direct real-robot deployment and supporting simulation and hardware results. The body supplied under this title and arXiv id is instead an unrelated condensed-matter manuscript on a Floquet–Keldysh functional renormalization group for the driven single-impurity Anderson model (frequency-dependent two-particle vertex, benchmarks vs 2PT/GW, Kondo resonance and photo-induced transport). No robotics methods, rewards, curricula, baselines, metrics, or robot experiments appear in the manuscript body.","tokens_in":28844,"tokens_out":787,"duration_ms":15147,"significance":"If the SmoothTurn claims were substantiated, the work would be of clear interest for high-speed legged navigation under sequential goals. As submitted, those claims cannot be assessed: the manuscript body does not contain the stated framework, experiments, or results. The Floquet-FRG content that is present may be significant in its own field, but it is not the paper under review and cannot support the robotics claims.","major_comments":[{"comment":"Title/abstract vs body mismatch: the abstract and paper_id describe SmoothTurn (sequential local navigation, sequential goal-reaching reward, lookahead goals, automatic goal curriculum, real quadruped deployment). The full text is a Floquet-Keldysh FRG study of the driven SIAM (Secs. I–V, Eqs. (1)–(79), Appendices A–C). There is no methods section, reward definition, observation design, curriculum, simulation protocol, baseline comparison, or real-robot experiment for SmoothTurn. The central empirical claim is therefore uncheckable from the submitted manuscript.","section":null},{"comment":"Load-bearing claim of sim and real-world success: the abstract asserts that SmoothTurn learns smooth turning with emergent momentum control, advance facing, and efficient paths, and deploys with onboard sensing/compute. No figures, tables, metrics, ablations of sequential reward / lookahead / curriculum, or hardware protocol exist in the provided body. Without that evidence the claim cannot be verified or revised within this document.","section":null},{"comment":"Weakest methodological assumption cannot be tested: the abstract frames single-goal policies as failing on sequential direction changes for lack of anticipation and momentum, and presents the three SmoothTurn ingredients as the remedy. The manuscript contains no sequential-navigation task definition, no comparison to single-goal policies, and no ablation of those ingredients, so the assumption remains unexamined.","section":null}],"minor_comments":[{"comment":"arXiv identifiers conflict: the cache labels paper_id 2603.12842 (SmoothTurn) while the body header cites arXiv:2603.12844 (Floquet FRG). The submission package appears corrupted or misassembled.","section":null},{"comment":"If the intended robotics manuscript is resubmitted, it should include explicit reward equations, lookahead window length, curriculum schedule, quantitative baselines/ablations, and a real-robot protocol with sensors and compute stack.","section":null}],"recommendation":"reject","confidential_remarks":"The PDF/body under review is the wrong paper relative to the title, abstract, and arXiv id. This looks like a pipeline or upload error rather than a scientific dispute. I recommend desk rejection or return without review until the correct SmoothTurn manuscript is provided; I have not evaluated the Floquet-FRG work as a robotics submission."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The arXiv id and abstract describe SmoothTurn: a sequential local-navigation formulation for high-speed multi-goal turns on quadrupeds, built from a sequential goal-reaching reward, future-goal lookahead, and an automatic goal curriculum, with claimed sim and onboard real-robot results and emergent momentum control, advance facing, and path efficiency.\n\nWhat is actually new, on the abstract’s own terms, is the task framing (sequential rather than single-goal local navigation) and the packaging of those three standard RL ingredients for smooth high-speed direction changes. That is a useful applied problem for rescue/inspection-style agility. The abstract is clear about the failure mode it targets—single-goal policies that park at the target and lose momentum across goal switches—and about direct deployment with onboard sensing and compute.\n\nThe soft spot is not subtle: the CACHEABLE full manuscript is not SmoothTurn at all. It is a Floquet–Keldysh FRG paper on the driven Anderson impurity (2603.12844). So there are no methods, baselines, ablations of reward/lookahead/curriculum, metrics, hardware protocol, or figures to check. The reader’s and stress-test notes are right: the strongest claim rests entirely on the abstract’s assertion of empirical success. Free parameters (reward weights, lookahead length, curriculum thresholds) and ordinary ML circularity risk cannot be assessed. Novelty and significance look mid-range if the claims hold; soundness is uncheckable from what we were given.\n\nWho this is for: people already working on learned quadruped loco-navigation who care about multi-goal agility. Without the correct PDF, I would not put it in reading group or cite it yet. A serious robotics venue would still send a coherent abstract of this type to referees once the right manuscript is attached—so yes on peer-review eligibility in principle—but right now we cannot judge the work itself. Get the real paper before spending more time.","headline":"We only have the SmoothTurn abstract; the supplied full text is a different paper, so the real-robot and emergent-behavior claims cannot be audited.","tokens_in":29477,"tokens_out":492,"would_cite":false,"duration_ms":10017,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A learning framework that makes quadrupedal robots turn smoothly at high speed by training on sequences of goals, not single targets.","keywords":["quadrupedal robots","agile navigation","sequential local navigation","smooth turning","reinforcement learning","goal curriculum","locomotion control","sim-to-real"],"falsifier":"Run the same robot and hardware stack with a single-goal baseline versus SmoothTurn on matched high-speed multi-waypoint courses; if SmoothTurn does not show measurably smoother turns, higher sustained speed through goal switches, or reliable onboard real-robot success, the central claim fails.","tokens_in":29563,"feed_emoji":"🐕","tokens_out":817,"duration_ms":13235,"temperature":0.7,"pith_summary":"High-speed rescue and inspection missions need robots that can change direction without losing momentum. Most learned controllers only train for one goal at a time and then sit still once they arrive, so they cannot plan the next turn or keep running through a series of waypoints. This paper reframes the job as sequential local navigation and introduces SmoothTurn: a reward that scores progress across a chain of goals, observations that include a short window of upcoming goals, and a curriculum that hardens the goal sequences only as the robot gets better. The resulting policy is claimed to run and turn fluidly, with spontaneous behaviors such as bleeding speed into a turn, aiming the body at the next goal early, and taking efficient paths—and it is said to transfer straight onto real quadrupeds using only onboard sensing and compute.","feed_headline":"Robots learn to turn smoothly while running fast","feed_subtitle":"Sequential goals, lookahead, and a curriculum unlock high-speed multi-waypoint quadruped turns","key_machinery":"SmoothTurn: a learning-based control stack whose three parts—a sequential goal-reaching reward, an observation space expanded by a future-goal lookahead window, and an automatic goal curriculum that grows sequence difficulty with performance—together force the policy to anticipate turns instead of treating each target as a stop.","core_discovery":"When a quadruped is trained with a sequential goal-reaching reward, a lookahead over future goals, and an automatic curriculum on goal-sequence difficulty, it learns an agile high-speed turning policy for sequential local navigation that maintains momentum across goal switches and can be deployed on real robots with onboard sensors and computation.","pith_inferences":["The same sequential-goal plus curriculum recipe may transfer to other platforms that suffer stop-and-turn behavior, such as bipeds or wheeled robots with tight turning constraints.","If lookahead length is a first-order knob, short horizons may still fail in cluttered or partially observed maps where the next goals are uncertain.","Real deployments will likely need the curriculum and reward to stay robust when goal sequences are generated online by a higher-level planner rather than sampled offline."],"forward_implications":["Sequential local navigation becomes a practical training objective for agile quadrupeds instead of stitching single-goal policies.","Policies can learn to bleed momentum, pre-orient the body, and cut efficient paths without hand-scripted turn controllers.","Trained controllers can be dropped onto real quadrupeds with only onboard sensing and compute for multi-waypoint missions.","Urgency-driven tasks such as fire rescue and industrial inspection gain a concrete path to high-speed direction changes."],"fun_headline_variants":["SmoothTurn teaches quadrupeds high-speed smooth sequential turns","Sequential goals plus lookahead yield agile quadruped turning","Curriculum-trained policy keeps momentum across goal switches","Quadrupeds learn smooth turns while running with future-goal lookahead","Real-robot deployable policy for agile multi-waypoint quadruped runs"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The main reason single-goal policies fail at fast sequential turns is that they cannot see the next goals or keep momentum, and adding those three training ingredients is enough to unlock smooth high-speed turning and real-robot transfer.","fun_headline_variants_meta":{"raw":{"variants":["SmoothTurn teaches quadrupeds high-speed smooth sequential turns","Sequential goals plus lookahead yield agile quadruped turning","Curriculum-trained policy keeps momentum across goal switches","Quadrupeds learn smooth turns while running with future-goal lookahead","Real-robot deployable policy for agile multi-waypoint quadruped runs"]},"model":"grok-4.5","effort":"low","cost_usd":0.00482,"raw_usage":{"total_tokens":1364,"prompt_tokens":793,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":48200000,"prompt_tokens_details":{"text_tokens":793,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":487,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":793,"tokens_out":84,"duration_ms":4639,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T22:04:38.617496+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the same robot and hardware stack with a single-goal baseline versus SmoothTurn on matched high-speed multi-waypoint courses; if SmoothTurn does not show measurably smoother turns, higher sustained speed through goal switches, or reliable onboard real-robot success, the central claim fails.","supporting_citations":[],"review_version":1}