{"id":"046753ef-c864-480f-9530-c1b23dc38b8b","arxiv_id":"2508.12928","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A Monte Carlo tree search and trajectory optimization pipeline that simultaneously chooses contact sequences and contact patches for legged robots, demonstrated on a quadruped and a humanoid.","lead":"This robotics paper combines Monte Carlo tree search with whole-body trajectory optimization to plan both the sequence of footholds and their exact positions on constrained terrain for legged robots. The authors show the planner finds diverse collision-free plans in simulation and successfully executes one on a real quadruped robot.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Feasibility of reported 'dynamically consistent, collision-free' plans rests on permissive proxies: KKT residuals without threshold and MuJoCo collision checks at N=50 nodes only, with no physics-rollout validation over the plan set.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point: the feasibility and collision-free properties of the generated plans are asserted on the basis of optimization residuals and discretized collision counts, not on independent verification. I agree. The reason this is the most load-bearing concern, rather than the missing baselines or the single hardware demo, is that the paper's central contribution is not the MCTS algorithm per se but the claim that the pipeline can quickly filter contact plans into a diverse set of dynamically consistent, collision-free trajectories. If the filter is permissive, the number of 'collision-free trajectories' in Fig. 6 is an overestimate and the whole-body TO + MCTS integration is not demonstrated to produce sound plans. The paper's own caveats in Section III-A.4 (no guarantees) and Section IV-D (non-negligible residuals for some humanoid plans) are internal evidence that the proxy is not airtight. The proposed test—closed-loop physics simulation of all reported plans—would settle whether the proxy is adequate. If it passes, the conditional verdict can be upgraded; if it fails, the claims need to be weakened. Since the reader already reached CONDITIONAL on this basis, no verdict change is needed, though the test would materially strengthen the paper.","tokens_in":9921,"tokens_out":6052,"duration_ms":64134,"concrete_test":"Run each of the plans counted in Fig. 6 (e.g., all successful runs across gap lengths and box heights) through the same closed-loop MuJoCo MPC used in Section IV-C, with the identical robot model and task terrain, and record the success rate (reaching the goal region without falling and without collision). Also re-sample each planned trajectory at 1 kHz and count MuJoCo collisions to detect collisions occurring between the N=50 nodes. If the large majority of plans succeed and no inter-node collisions appear, the concern is resolved; if a substantial fraction fail, the 'dynamically consistent, collision-free' claim should be restricted to the subset verified by hardware/rollout.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the pipeline 'quickly finds a diverse set of dynamically consistent, collision-free plans' is not directly validated. The MCTS reward uses R_res, computed from KKT residuals of a non-convex trajectory optimization (Eq. 3), with no reported threshold for what counts as feasible; the paper explicitly states zero constraint violations cannot be guaranteed (Section III-A.4). The collision-free label rests on MuJoCo collision counts at N=50 discretized trajectory nodes, so collisions between nodes are systematically unobserved. Figure 6 reports counts of 'collision-free trajectories' without stating a residual cutoff or success rate in a physics rollout. The single hardware demo (0.18 m box) and the humanoid section's admission that 'constraint residuals in some of the motions are not negligible' (Section IV-D) further indicate that the feasibility proxy is not a hard guarantee. If the proxy is too permissive, the main contribution—fast generation of dynamically consistent, collision-free plans—is unsupported for the broad parameter sweep in Fig. 6, and the paper's strongest claim reduces to 'plans that score well under a permissive proxy.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a pipeline that combines Monte-Carlo Tree Search (MCTS) with whole-body trajectory optimization (TO) to plan contact sequences and contact patches simultaneously for legged robots. The MDP formulation encodes contact modes and patch assignments, graph pruning reduces the search space, and candidate plans are scored by a reward that combines a collision count penalty with a sigmoid of the product of KKT residuals of the trajectory optimization. The TO component uses centroidal momentum dynamics, whole-body kinematics, foot contact constraints, and several end-effector level geometric costs. Experiments cover a Unitree Go2 quadruped crossing gaps up to 1.5 m and climbing boxes up to 0.7 m, with statistics over five runs, one hardware demonstration of a 0.18 m box climb, and a humanoid box-climbing example with 65 reported collision-free trajectories. The paper claims the first demonstration of simultaneous contact sequence and patch selection for acyclic multi-contact locomotion using the whole-body dynamics of a quadruped.","tokens_in":10294,"tokens_out":4200,"duration_ms":45322,"significance":"If the reported validation is tightened, this is a useful contribution to contact planning for legged robots. The pipeline integrates MCTS with whole-body TO in a way that goes beyond earlier centroidal or predefined-gait formulations, and the hardware demonstration, while limited, is a meaningful step beyond pure simulation. The paper also explicitly aims at generating diverse solution sets, which is interesting for downstream imitation-learning applications. However, the central feasibility claims are currently supported by proxies: KKT residuals without a stated acceptance threshold and collision checks at a fixed number of discrete trajectory nodes. The humanoid section itself concedes that some residuals are not negligible. These gaps affect the strongest claims in the abstract and conclusion, so the manuscript needs additional validation before the main claims can be accepted as stated.","major_comments":[{"comment":"The reward R_res uses the product of KKT residuals through a sigmoid, but no acceptance threshold or actual residual values are reported. Section III-A.4 admits that zero constraint violations cannot be guaranteed, and Section IV-D states that some humanoid motions have non-negligible residuals. As a result, the counts of 'collision-free trajectories' in Fig. 6 and the abstract's 'dynamically consistent plans' rest on an unquantified feasibility proxy. Please report the residual distributions, specify the threshold used to accept a plan, and verify a random subset of accepted plans by physics rollout.","section":"Section III-A.4, Eq. (3)"},{"comment":"Collision avoidance is evaluated by counting collisions at the N=50 discretized trajectory nodes in MuJoCo. A fast-moving link can pass through an obstacle between nodes, so this discretization does not establish 'collision-free' in the sense claimed for Fig. 6. Please add continuous collision checking, a denser time grid, or dynamic simulation of the complete plan, and report the resulting success rate over the accepted plans.","section":"Section III-A.4 and Section IV-B"},{"comment":"The hardware transfer demonstration is a single quadruped climb onto a 0.18 m box, whereas the simulation sweep in Fig. 6 claims plans for box heights up to 0.7 m and gap lengths up to 1.5 m. One easy case does not establish that the pipeline's plans are generally transferable. Please specify the selection procedure for the hardware plan, report additional hardware trials or at least state this as a limitation, and avoid overgeneralizing 'transferable' to the full parameter range.","section":"Section IV-C, Fig. 1"},{"comment":"The humanoid results state that 'dynamic feasibility of the trajectories is satisfied for most cases, but not all' and that 'constraint residuals in some of the motions are not negligible.' This directly contradicts a blanket 'dynamically consistent' claim if applied to the humanoid experiments. The abstract and conclusion should be scoped to the accepted quadruped plans, and the residual metric used to decide 'satisfied' should be defined for the humanoid, with the non-negligible cases excluded from the reported counts.","section":"Section IV-D"}],"minor_comments":[{"comment":"Please clarify whether the five runs use different random seeds and specify how the standard-deviation lines are computed; the axis labels in Figs. 6 and 7 would also benefit from explicit units.","section":"Section IV-B, Figs. 6 and 7"},{"comment":"The text states that the reference patch changes halfway through the swing phase, but it does not explain how this switch is encoded in the trajectory optimization, which may be piecewise or require additional decision variables.","section":"Section III-B.2"},{"comment":"The relationship between the parameter beta_plus and the resulting swing height is only intuitive; a short derivation or explicit dependence would make the cost easier to interpret.","section":"Eq. (11a) and Fig. 4"},{"comment":"Please clarify the relation between '300 MCTS iterations' and the reported 'first collision-free solution after 52 seconds', since the run time and iteration count are not obviously linked in the text.","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"I see no serious novelty or citation concerns: the paper builds on the authors' prior MCTS+MPC work and states its claims in relation to that work. The main issue is the gap between the permissive feasibility proxies and the strong 'dynamically consistent, collision-free' wording in the abstract and conclusion. If the authors add residual thresholds and physics-rollout validation, I would view the paper as a solid systems contribution; without that, the central claim remains unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real: it combines MCTS with whole-body trajectory optimization to plan contact sequences and patches at the same time, using full nonlinear dynamics of a quadruped. Prior work either used simplified dynamics with no patch selection [19, 20] or pre-defined gaits [21, 22]. This combination is a legitimate step forward, and the results show it: diverse plans found in seconds for gap crossing and box climbing up to 1.5 m and 0.7 m, plus acyclic humanoid maneuvers that are genuinely hard for gradient-based methods. The formulation is coherent. The graph pruning (minimum contact duration, no patch change without breaking, no leg crossing) is sensible, and the swing-phase costs with the switching reference frame are thoughtful. The authors are also honest about limitations: they state zero constraint violations cannot be guaranteed, and they admit the humanoid residuals are non-negligible in some motions. The soft spots are in the gap between the claims and the evidence. The 'dynamically consistent, collision-free' label rests on KKT residuals without a reported threshold and on MuJoCo collision counts at N=50 discretized nodes. Collisions between nodes are not checked, and there is no physics rollout validation over the plan set. The single hardware demo (a 0.18 m box climb) is a real external test, but it lacks quantitative tracking data, so it does not fully validate the broad parameter sweep in Figure 6. That concern is proportionate: the pipeline's output is not fake, but the feasibility claims are supported by proxies, not guarantees. Two more weaknesses matter. There are no baseline comparisons, so the 'quickly finds' claim is hard to contextualize. The hyperparameters are numerous (reward scalings, UCB constant, swing cost parameters, patch safety margins) with no sensitivity analysis, and five runs per task is thin for the statistics reported. No code or data is released, which hurts reproducibility for a planning paper. None of this is fatal. The central mechanism is sound, the hardware transfer is a genuine external check, and the residual-based reward is not circular in a harmful way because the final plans are executed. The paper is for anyone working on contact-implicit planning, MCTS for locomotion, or dataset generation for learning-based policies. It deserves a serious referee. I would ask for baseline comparisons, a physics-rollout validation or explicit residual thresholds, and ideally code release, but this is a solid contribution that should be reviewed, not desk-rejected.","headline":"A genuinely useful MCTS + whole-body TO pipeline for simultaneous contact sequence and patch selection, with real hardware transfer, but the feasibility claims are softer than the abstract suggests.","tokens_in":772,"tokens_out":729,"would_cite":true,"duration_ms":25706,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A tree search plus whole-body optimization selects contact sequence, landing patches, and motion in one pass.","keywords":["contact planning","legged locomotion","Monte-Carlo tree search","whole-body trajectory optimization","contact sequence selection","contact patch selection","acyclic multi-contact locomotion","quadruped and humanoid"],"falsifier":"Take a plan the pipeline labels collision-free and simulate it in a high-fidelity physics engine with dense time stepping and continuous collision detection, then check whether any link penetrates the environment or the friction cone is violated between the 50 sampled nodes. A single reported plan that fails this check would show the collision and feasibility proxies are permissive; running the same check across the humanoid set would settle how strong the claim is.","tokens_in":9749,"feed_emoji":"🦿","tokens_out":8540,"duration_ms":86840,"temperature":0.7,"pith_summary":"Legged locomotion planning usually separates deciding which surfaces to touch and in what order from computing the motion itself, or simplifies the dynamics to make that combined problem tractable. This paper claims that a pipeline combining Monte-Carlo tree search with whole-body trajectory optimization can do both at once: the search proposes contact sequences and landing patches, and the trajectory optimizer checks each candidate against full nonlinear dynamics. For a quadruped crossing a gap or climbing onto a box, the pipeline finds several distinct plans that pass its dynamics and collision checks within seconds, and one box-climbing plan is executed on real hardware under a 30 Hz model-predictive controller. The same pipeline also produces acyclic humanoid maneuvers such as climbing a 0.5 m box. If the claim holds, this closes a gap between discrete contact planning and dynamic motion planning that earlier mixed-integer or gait-fixed methods left open.","feed_headline":"Planner chooses contact sequence, landing patches, and motion at once","feed_subtitle":"The search-plus-optimization pipeline finds diverse collision-free quadruped plans in seconds and transfers one to hardware.","key_machinery":"The mechanism is a search-then-verify loop. Monte-Carlo tree search, a tree-search algorithm that balances exploration and exploitation, explores a graph whose nodes are contact modes—for each end-effector, whether it is on the ground and which numbered patch it uses—and whose edges switch feet between keeping, breaking, and making contact, with pruning for minimum contact duration, patch changes without breaking contact, and leg crossing. Each candidate full plan is passed to a whole-body trajectory optimization that uses centroidal momentum dynamics and full kinematics and optimizes states, velocities, and contact forces over N=50 nodes, enforcing non-sliding contact, friction cones, foot-above-surface constraints, patch-boundary limits, and swing-phase collision-avoidance costs. The search reward is the product of a collision term computed from collisions counted at the 50 nodes and a residual term derived from the optimizer's optimality-condition residuals, so approximate dynamic feasibility and collision avoidance are the guiding signals for the discrete search.","core_discovery":"The central claim is that discrete contact decisions and continuous motion can be planned simultaneously without dropping to linear or centroidal-only dynamics. Each complete contact plan—which end-effectors are in contact at each phase and on which rectangular patch—is scored by running a whole-body trajectory optimization; the optimizer's residuals and the number of collisions counted on the discretized trajectory define the tree-search reward. The paper reports that this identifies many dynamically consistent, collision-free plans for both a gap-crossing task and a box-climbing task across a range of difficulty, with first solutions typically under 10 seconds on easier cases. A planned box climb executed on a real quadruped in a model-predictive control loop demonstrates transfer, and a humanoid variant finds 65 collision-free box-climbing trajectories. The paper states the transferability result as the first demonstration of simultaneous contact sequence and patch selection for acyclic multi-contact locomotion using the whole-body dynamics of a quadruped, while noting that some humanoid trajectories retain non-negligible constraint residuals.","pith_inferences":["The reported per-optimization time and seconds-to-minutes planning times mean the current pipeline is a batch planner, not an online one; a natural extension, which the paper names as future work, is to distill the found plans into a learned policy that selects contact sequences at runtime.","Collision checks at 50 discrete nodes leave continuous-time penetration possible between nodes; testing the same reward with swept-volume collision detection or much finer time stepping would turn 'collision-free' into a stronger statement.","The reward's reliance on optimizer residuals makes the planner's notion of dynamic feasibility depend on optimizer convergence; as the paper itself notes for some humanoid trajectories, adding a separate post-hoc verification step would harden the pipeline.","The same search-then-optimize recipe should transfer to whole-body manipulation, where the end-effectors are hands and patches are graspable surfaces, since the formulation only assumes end-effector contact modes and patch geometry."],"forward_implications":["Because contact modes are generated rather than chosen from a gait library, the same planner can produce acyclic, non-repeating sequences such as using a wall as an intermediate support or climbing onto a box.","First solutions arrive within seconds for easier parameter ranges, and several collision-free alternatives per minute are found, so the pipeline can serve as a dataset generator for imitation learning of contact-planning policies.","The formulation extends across robot morphologies with minimal changes, so contact-planning search developed on a quadruped transfers to a humanoid with only modified contact models.","Because each candidate is verified by whole-body optimization, the final plans are expressed as full state-and-force trajectories ready to be tracked by a model-predictive controller, as demonstrated at 30 Hz replanning on hardware."],"supporting_citations":[{"why":"Supplies the whole-body trajectory optimization scheme based on centroidal dynamics and full kinematics that the paper adapts for per-plan verification.","marker":"[24]"},{"why":"Provides the MCTS-based non-gaited contact-sequence planner that this work extends to simultaneous patch selection with full dynamics.","marker":"[19]"},{"why":"Baseline that combines MCTS with nonlinear model-predictive control on constrained surfaces but relies on pre-defined gaits, which this work removes.","marker":"[21]"},{"why":"Related MCTS-plus-model-predictive-control contact planner with learned transition feasibility, framed as a baseline for patch selection.","marker":"[22]"},{"why":"Prior contact-explicit mixed-integer convex optimization that plans contact, gait, and motion but requires simplified dynamics.","marker":"[14]"},{"why":"Prior efficient multi-contact pattern generation via sequential convex approximation of centroidal dynamics, the simplified-dynamics baseline being generalized.","marker":"[15]"},{"why":"Supplies the upper-confidence-bound rule used for node selection in the Monte-Carlo tree search.","marker":"[23]"},{"why":"Provides the physics and visualization engine used to count collisions on discretized trajectories and to test execution in simulation.","marker":"[28]"},{"why":"Supplies the sequential quadratic programming solver used to run the trajectory optimizations and the real-time model-predictive controller.","marker":"[25]"}],"fun_headline_variants":["Joint contact and patch planning via Monte-Carlo tree search and whole-body optimization","Combined contact sequence and patch selection for dynamic legged locomotion","Simultaneous contact and patch decisions with whole-body trajectory optimization","Tree search plus whole-body optimization yields diverse legged plans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a plan is accepted as dynamically reliable if the trajectory optimizer reports that its equations are nearly satisfied and if no collision is spotted at the 50 sampled instants; if either check is too lenient, the claimed 'dynamically consistent' and 'collision-free' plans may not actually be feasible.","fun_headline_variants_meta":{"raw":{"variants":["Joint contact and patch planning via Monte-Carlo tree search and whole-body optimization","Combined contact sequence and patch selection for dynamic legged locomotion","Simultaneous contact and patch decisions with whole-body trajectory optimization","Tree search plus whole-body optimization yields diverse legged plans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1332,"prompt_tokens":885,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":374}},"tokens_in":501,"tokens_out":447,"duration_ms":5067,"temperature":1.0,"reasoning_tokens":374,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:16:56.162767+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a plan the pipeline labels collision-free and simulate it in a high-fidelity physics engine with dense time stepping and continuous collision detection, then check whether any link penetrates the environment or the friction cone is violated between the 50 sampled nodes. A single reported plan that fails this check would show the collision and feasibility proxies are permissive; running the same check across the humanoid set would settle how strong the claim is.","supporting_citations":[{"cited_title":"Monte carlo tree search gait planner for non-gaited legged system control,","cited_arxiv_id":null,"evidence_quote":"Provides the MCTS-based non-gaited contact-sequence planner that this work extends to simultaneous patch selection with full dynamics."},{"cited_title":"Diffusion-based learning of contact plans for agile locomotion,","cited_arxiv_id":null,"evidence_quote":"Baseline that combines MCTS with nonlinear model-predictive control on constrained surfaces but relies on pre-defined gaits, which this work removes."},{"cited_title":"Learning feasible transitions for efficient contact planning","cited_arxiv_id":"2407.11788","evidence_quote":"Related MCTS-plus-model-predictive-control contact planner with learned transition feasibility, framed as a baseline for patch selection."},{"cited_title":"Simultaneous contact, gait, and motion planning for robust multilegged locomotion via mixed-integer convex optimization,","cited_arxiv_id":null,"evidence_quote":"Prior contact-explicit mixed-integer convex optimization that plans contact, gait, and motion but requires simplified dynamics."},{"cited_title":"Efficient multicon- tact pattern generation with sequential convex approximations of the centroidal dynamics,","cited_arxiv_id":null,"evidence_quote":"Prior efficient multi-contact pattern generation via sequential convex approximation of centroidal dynamics, the simplified-dynamics baseline being generalized."},{"cited_title":"Bandit based monte-carlo planning,","cited_arxiv_id":null,"evidence_quote":"Supplies the upper-confidence-bound rule used for node selection in the Monte-Carlo tree search."},{"cited_title":"Mujoco: A physics engine for model-based control,","cited_arxiv_id":null,"evidence_quote":"Provides the physics and visualization engine used to count collisions on discretized trajectories and to test execution in simulation."}],"review_version":1}