{"id":"ec1bf35d-ef1d-48c3-891b-99d72cf1e918","arxiv_id":"2506.09979","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A layered controller that samples footholds and runs parallel fixed-mode MPC evaluates terrain options in real time, enabling a quadruped and a simulated humanoid to traverse stepping stones.","lead":"Legged robots usually either fix their footholds in advance or use slow mixed-integer optimization to choose them. This paper describes a two-layer controller that samples possible footholds and evaluates each with a fast optimal controller in parallel, then executes the cheapest plan in real time, demonstrated on a quadruped and a simulated humanoid.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reliability and optimality comparison is not yet established: 5 trials per method and an MPC-cost metric that the sampler directly minimizes; fixed contact schedule is a real but secondary limitation.","rationale":"The paper's main technical contribution is a sensible layered architecture with a real hardware demonstration and plausible real-time timings, so I do not see a reason to reject it. However, the strongest abstract claim is comparative, and that comparison is the least secure part. The experiment in Sec. IV uses only five trials per method; with such small n, the observed failure difference is within the range of chance. The optimality comparison is also confounded by using the controller's own cost as the metric while the sampling layer explicitly selects the lowest-cost MPC solve; the fact that the cost includes a Raibert heuristic term makes the comparison with a Raibert baseline particularly hard to interpret. The fixed contact schedule identified by the reader is a genuine limitation and is acknowledged in the paper and its conclusion, but it applies equally to both compared methods in the experiment, so it is not the main threat to the central comparative claim. A direct timing comparison against pure sampling methods is also absent, but that claim is less central and the reported per-solve timings are at least plausible. Given the hardware result and the architectural novelty, the appropriate verdict remains conditional: accept the feasibility contribution, but require stronger, metric-appropriate comparison before accepting the optimality/reliability claim at face value.","tokens_in":11558,"tokens_out":4721,"duration_ms":56793,"concrete_test":"Reproduce the Go2 stepping-stone comparison with at least 20 randomized trials per method, recording success/failure and task-level metrics independent of the MPC objective (e.g., time to reach the final platform, total mechanical energy, torso path deviation from the centerline) in addition to the MPC cost. Also rerun the comparison with the Raibert end-effector term removed from the MPC cost. If the sampling method does not show a statistically reliable improvement on the task-level metrics, the \"more optimal and reliable\" claim should be downgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparative claim (abstract: \"more optimal and reliable than common heuristic-based approaches\") rests on the simulation experiment in Sec. IV. Two features make that evidence weak. First, the sample size is five runs per method; the reported 80% vs 20% failure difference (Raibert fails 4/5, sampling fails 1/5) is not statistically meaningful. Second, the optimality metric is the same MPC cost that the sampling layer minimizes: each sample is an MPC solve and the lowest-cost trajectory is selected, so a lower average cost for sampling is expected by construction. The cost also includes a Raibert end-effector placement term (Sec. III), i.e., the very heuristic used by the baseline, further muddying the comparison. Reporting average cost only across successes (one heuristic success vs four sampling successes) makes the comparison still weaker. The hardware demonstration supports feasibility and real-time operation, and the fixed contact schedule (Sec. II-A) is a stated scope limitation, but the abstract's optimality/reliability claim needs stronger evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a layered controller for legged locomotion over constrained footholds. A gradient-free sampling layer selects which terrain polytopes to use, and a fixed-mode whole-body MPC is solved in parallel for each sampled option; the lowest-cost MPC solution is then executed. The fixed-mode MPC uses a quaternion tangent-space parameterization, inverse-dynamics constraints, polytope foot-placement constraints, and a Raibert-style end-effector cost term. Results include hardware experiments on a Unitree Go2 quadruped and asynchronous MuJoCo simulations of a Unitree G1 humanoid. The authors claim that the layered approach is more optimal and reliable than heuristic terrain selection and faster to compute than pure sampling methods.","tokens_in":11800,"tokens_out":4227,"duration_ms":58094,"significance":"If fully supported, the paper would be a useful practical contribution: the two-layer decomposition is clean, the hardware demonstration on a real quadruped is encouraging, and the real-time humanoid simulation suggests the architecture scales to higher-dimensional systems. The technical formulation is largely standard but competently presented, and the detailed timing breakdown is helpful. The paper does not provide code or formal proofs, and the central comparative claims currently rest on weak evidence; the feasibility result is the strongest part of the paper, while the optimality and reliability claims are not yet established.","major_comments":[{"comment":"The reliability comparison is based on only five simulation runs per method, with the Raibert heuristic failing 80% of the time and the sampling method failing once. With n=5, this difference is not statistically meaningful; a Fisher exact test would not reach conventional significance. The statement that the sampling method 'outperforms the Raibert heuristic by a significant margin' is therefore unsupported. The authors should run more trials, report confidence intervals or a statistical test, or weaken the claim to an anecdotal observation.","section":"Sec. IV, reliability experiment"},{"comment":"The optimality comparison is largely circular. The sampling layer selects the lowest-cost MPC solve, and the reported cost is exactly the MPC cost in Eq. (8), so lower average cost for the sampling method is expected by construction. In addition, Sec. III states that the cost function includes the Raibert end-effector placement term, which is the very heuristic used by the baseline; with only one heuristic success versus four sampling successes, averaging the cost over successes introduces further selection bias. An external performance metric (for example, success rate with a larger number of trials, distance traveled, or time to completion) or a cost ablation without the Raibert term is needed.","section":"Sec. IV and Eq. (8)"},{"comment":"The claim that the method is 'faster to compute than pure sampling methods' is not supported by direct comparison. Table I reports only the proposed algorithm's own timings on a single CPU, while the comparisons to CIMPC and to sampling-based methods rely on cited timings from different implementations and different hardware. A head-to-head benchmark under the same timing protocol, or at least a same-hardware reimplementation of a baseline, is required before this claim can be evaluated.","section":"Sec. IV and Table I"}],"minor_comments":[{"comment":"The sentence 'This will effect how the constraints are formulated' should read 'affect'.","section":"Sec. II-A"},{"comment":"The phrase 'Simulation results shown the efficacy' should read 'show the efficacy'.","section":"Sec. I"},{"comment":"The kinematic-limit circle radius and the polytope shrink amount are not quantified; these are additional free parameters that should be listed explicitly in the experimental setup.","section":"Sec. II-B"},{"comment":"The description of the humanoid actuation (19 of 27 actuators optimized, with ankle and wrist joints fixed) appears only in the implementation section; stating the resulting degrees of freedom and how the fixed joints are held would improve reproducibility.","section":"Sec. III"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems paper with a clear feasibility demonstration, but the headline comparative claims are stronger than the evidence. I would ask for either more extensive comparisons or a softening of the abstract's optimality/reliability/speed claims. The fixed contact schedule and single-SQP-iteration choices are acceptable engineering compromises if framed as such rather than as general optimality guarantees."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zach,\n\nHere's my read of Olkin and Ames, the layered foothold-selection MPC.\n\nWhat's actually new: they take their earlier layered MPC framework and apply it to terrain selection by sampling foothold candidates, solving a fixed-mode full-order MPC in parallel for each sample, and executing the lowest-cost trajectory. That combination — sampling discrete choices, full-order MPC for each, real-time on hardware — is new for this application. The quaternion tangent-space state parameterization is a minor but sensible improvement. The hardware demonstration on the Go2 stepping over cinder blocks and gaps is real evidence that the real-time claim holds; the humanoid simulation also runs in real time. The implementation details, including the preparation/feedback split and interpolation, are described carefully enough to reproduce.\n\nThe paper's main weakness is the comparative evidence. The abstract claims the method is \"more optimal and reliable\" than heuristics and faster than pure sampling, but the reliability comparison rests on five simulation runs per method: Raibert fails 4/5, sampling fails 1/5. That difference is not statistically meaningful. The optimality metric is the MPC cost in Eq. 8 — the same quantity the sampling layer minimizes when it selects the lowest-cost solve. So lower average cost for sampling is partly by construction. It doesn't help that the cost includes a Raibert end-effector term, the same heuristic used as the baseline. And reporting average cost only across successes (one heuristic success vs. four sampling successes) makes the comparison weaker, not stronger. The timing claims are also not based on direct head-to-head benchmarks; they compare against published timings from other papers, which is fine as a ballpark but doesn't support a strong \"faster than pure sampling\" statement.\n\nThe fixed contact schedule is a stated limitation (Sec. II-A), and it's real — if terrain forces an asymmetric gait, the method can't adapt the timing. But it's a scope limitation, not a hidden flaw. The paper is honest about it.\n\nOverall: the architecture is plausible, the hardware demo supports the feasibility and real-time claims, and the paper is a useful engineering contribution. The comparative optimality/reliability claims need a bigger experiment and a non-circular metric before they should be taken at face value. I'd send it to peer review — it deserves referee time — but the revision should either strengthen the comparison or soften the abstract.\n\nWould cite if I were working on legged MPC. Worth a reading group session, mostly to discuss how to fairly evaluate sampling-based discrete decisions.\n\nBest.","headline":"Solid layered MPC architecture with a genuine hardware demo; the optimality/reliability comparison is underpowered and partly circular.","tokens_in":12289,"tokens_out":1878,"would_cite":true,"duration_ms":22865,"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 layered architecture that separates foothold sampling from whole-body MPC lets legged robots traverse stepping-stone terrain in real time, with better cost and reliability than a conventional foot-placement heuristic.","keywords":["legged locomotion","model predictive control","layered control architecture","terrain sampling","foothold selection","quadruped robot","humanoid robot","real-time control"],"falsifier":"Run the same controller on a sparse stepping-stone course while sweeping the number of parallel samples M; if success rate stops improving before worst-case terrain is handled, the remaining failures come from the fixed contact schedule rather than from insufficient terrain exploration.","tokens_in":11357,"feed_emoji":"🦿","tokens_out":10845,"duration_ms":118518,"temperature":0.7,"pith_summary":"This paper argues that real-time legged locomotion over constrained footholds does not need a single optimization over both discrete and continuous variables. Its claim is that a two-layer architecture—a sampling-based layer that proposes footholds, plus a full-order model predictive controller solved in parallel for each proposal—can pick terrain online while keeping control rates fast enough for hardware. On a quadruped crossing gaps and height-varying stones, the sampling layer failed once in five runs at 0.7 m/s whereas the conventional heuristic failed four times, and the successful sample-based runs had lower average cost. The same controller ran in real time on a simulated humanoid, which used its arms and waist to keep the torso upright while swinging legs over gaps. If correct, the architecture offers a general way to make hybrid locomotion decisions without hand-coded foot-selection rules.","feed_headline":"Sampling plus model-predictive control picks footholds in real time","feed_subtitle":"A layered controller beats conventional foot-placement heuristics and runs on a simulated humanoid.","key_machinery":"The load-bearing object is the fixed-mode MPC: a full-order nonlinear program over joint configuration, velocity, torque, and end-effector forces, in which the contact schedule is fixed and each candidate foothold enters as a polytope constraint on the stance foot, $\\gamma_k^l A_k^l r^l(q_k) \\le \\gamma_k^l b_k^l$. The discrete layer builds a sample tree around a nominal foot position projected from the velocity command; each branch of the tree is a terrain combination, sampled in proportion to the area of each terrain piece inside the robot's kinematic reach, and each branch is evaluated by one parallel MPC solve. The MPC uses a real-time iteration scheme—one quadratic program per control step—with inverse dynamics and quaternion tangent-space coordinates, so each evaluation is cheap enough to run several in parallel and still close the loop at 100 Hz. The point of the machinery is that discrete choices are scored by the actual closed-loop cost of the full-order controller rather than by a heuristic.","core_discovery":"The central claim is that hybrid locomotion decisions can be split by variable type without sacrificing either optimality or real-time feasibility. Discrete choices—which foothold polytope each foot will land on—are made by a gradient-free sampling layer, fixed as constraints in a smooth nonlinear MPC, and the lowest-cost MPC solution over all samples is executed. The paper reports that on stepping stones with varying heights at 0.7 m/s, the sampling controller failed once in five runs while the nominal-foot-placement heuristic failed in four; even the heuristic's successes had a higher average cost (1.0963 versus 0.8946). The same controller steered a simulated humanoid across gaps in real time with two samples, and the whole-body formulation let the robot use its arms and waist for stabilization. This is presented as evidence that the layered approach is more optimal and reliable than heuristic terrain selection and faster than pure sampling-based full-order optimization.","pith_inferences":["Editorial inference: the reliability advantage of sampling over the heuristic should grow with speed and terrain sparsity, because the nominal foot position is exactly the predictor that degrades in those regimes; the 0.7 m/s experiment is one point on that curve.","Editorial inference: the fixed contact schedule, not the sampling distribution, is the operative limit; a terrain that demands an atypical gait phase cannot be handled by any amount of foothold sampling, so generalizing the layer to contact timing is the natural next test.","Editorial inference: the same split of discrete sampling plus parallel gradient-based evaluation could transfer to manipulation, where the discrete options are grasp points or wall contacts and the continuous problem is a whole-body MPC."],"forward_implications":["With four parallel MPC evaluations, the hardware quadruped crossed the stepping-stone course at 0.7 m/s with one failure in five runs, versus four failures for the heuristic, and with lower average cost on successes.","Because terrain choice is scored by the MPC cost rather than a fixed rule, the same sampling layer can be pointed at other discrete decisions the paper names as future work, such as contact schedules or arm-bracing contacts.","The measured solve times—about 1.2 ms preparation and 4.3 ms feedback for four parallel quadruped samples, and 1.8 ms and 6.2 ms for two humanoid samples—fit inside a 10 ms control period, leaving room for more samples or longer horizons.","The controller works without a task-specific foothold heuristic; the nominal position is used only to center the sampling region, not to decide where to step."],"supporting_citations":[{"why":"Supplies the layered-architecture principle: separate discrete decisions from a stabilizing MPC layer.","marker":"[6]"},{"why":"Provides the polytope foot-placement constraint formulation used to impose the sampled terrain in the MPC.","marker":"[10]"},{"why":"Defines the nominal foot-placement heuristic (foot below the hip at mid-stance) used as the baseline.","marker":"[24]"},{"why":"Provides the real-time iteration scheme that reduces each MPC step to one quadratic program.","marker":"[25]"},{"why":"Supplies the high-performance quadratic-program solver used for every fixed-mode MPC solve.","marker":"[27]"},{"why":"Supplies rigid-body dynamics and analytical derivatives for the inverse-dynamics constraints.","marker":"[31]"},{"why":"Is the pure-sampling full-order MPC baseline the paper claims to be faster than.","marker":"[20]"},{"why":"Supplies contact-implicit MPC timing data used for speed comparison.","marker":"[11]"}],"fun_headline_variants":["Layer sampling with MPC for real-time legged footholds","Sampling plus MPC beats heuristics for step selection","Layered control: sample footholds, then MPC optimizes","Hybrid locomotion solved by splitting discrete and smooth","Real-time foothold choice via layered MPC and sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The contact schedule—which feet are in stance or swing at each instant—is fixed in advance and never sampled or optimized; only the terrain beneath those fixed foot timings is chosen.","fun_headline_variants_meta":{"raw":{"variants":["Layer sampling with MPC for real-time legged footholds","Sampling plus MPC beats heuristics for step selection","Layered control: sample footholds, then MPC optimizes","Hybrid locomotion solved by splitting discrete and smooth","Real-time foothold choice via layered MPC and sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000331,"raw_usage":{"total_tokens":1819,"prompt_tokens":895,"completion_tokens":924,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":846}},"tokens_in":511,"tokens_out":924,"duration_ms":7985,"temperature":1.0,"reasoning_tokens":846,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:35:44.024854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same controller on a sparse stepping-stone course while sweeping the number of parallel samples M; if success rate stops improving before worst-case terrain is handled, the remaining failures come from the fixed contact schedule rather than from insufficient terrain exploration.","supporting_citations":[{"cited_title":"Layered Nonlinear Model Predictive Control for Robust Stabilization of Hybrid Systems","cited_arxiv_id":"2503.12810","evidence_quote":"Supplies the layered-architecture principle: separate discrete decisions from a stabilizing MPC layer."},{"cited_title":"Raibert, Legged Robots That Balance","cited_arxiv_id":null,"evidence_quote":"Defines the nominal foot-placement heuristic (foot below the hip at mid-stance) used as the baseline."},{"cited_title":"A Real-Time Iteration Scheme for Nonlinear Optimization in Optimal Feedback Control,","cited_arxiv_id":null,"evidence_quote":"Provides the real-time iteration scheme that reduces each MPC step to one quadratic program."},{"cited_title":"The Pinocchio C++ library : A fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,","cited_arxiv_id":null,"evidence_quote":"Supplies rigid-body dynamics and analytical derivatives for the inverse-dynamics constraints."}],"review_version":1}