{"id":"f96a3c72-679f-4404-9f95-114b554485fc","arxiv_id":"2505.13715","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A real-time MPC controller with adjustable step timing and mixed-integer foothold selection lets bipedal robots step over or around ground obstacles while avoiding fast-moving body obstacles.","lead":"Bipedal robots can now plan their body path and foot placements together in real time, letting them dodge fast-moving obstacles and choose to step over or around objects on the floor. The framework was tested in simulations of Cassie and Digit robots and in a hardware walk-over experiment on a Digit humanoid.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Foot-avoidance guarantee hinges on an unverified tracking assumption: the MPC only constrains swing-foot endpoints and a straight-line height profile, and the claimed safety margin (0.4 m for a 0.2 m obstacle) is never checked against actual low-level tracking error.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the MPC's 3D foot-avoidance guarantee relies on a simplified swing-foot model with no dynamics or kinematics in the optimization. This is the single point where the central claim is most exposed. The paper's central claim is that it is the first model-based approach that \"implicitly chooses to step over or navigate around foot-level obstacles depending on CoM dynamics.\" For that claim to hold, the generated foot trajectory must actually avoid the obstacle in the physical system. Yet the optimization only constrains the start and end footholds in x-y and imposes height constraints along a straight line between them; the real trajectory is produced by a lower-level OSC that is never quantitatively evaluated for tracking error. The padding (0.4 m diameter for a 0.2 m obstacle) is a heuristic margin, not a verified bound. The authors acknowledge the omission by listing \"introducing swing foot kinematics/dynamics via nonlinear MPC\" as future work, which supports the interpretation that this is a known limitation rather than an intentional, validated design choice. Other concerns, such as lack of benchmarking against [28] and missing code/data, affect novelty and reproducibility but do not directly threaten the functional claim. A computational experiment that measures actual clearance over a range of conditions would settle whether the straight-line/tracking assumption is acceptable. Should it fail, the paper's central claim would need to be narrowed to the demonstrated cases; should it pass, the conditional verdict can be upgraded. Since the reader's CONDITIONAL verdict already captures this uncertainty, no adjustment is needed.","tokens_in":10338,"tokens_out":3922,"duration_ms":36553,"concrete_test":"Re-run the Digit foot-avoidance controller in simulation across a grid of obstacle lateral offsets (±0.3 m in 0.05 m steps) and commanded speeds (0.2–0.8 m/s), logging the full 3D swing-foot pose and the minimum clearance between the foot geometry and obstacle at 1 kHz. If the minimum clearance from the actual (simulated) foot trajectory falls below the assumed padding margin (0.1 m from the 0.4 m collision diameter) or contacts the obstacle in any trial, the straight-line height-constraint assumption is violated and the \"foot avoidance\" claim must be qualified. A complementary analytic check is to compute the worst-case OSC foot-tracking error from the leg Jacobian and compare it with the padding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. II-E.1 the x-y foot avoidance constrains only the starting and landing footholds to be in a chosen collision-free quadrant (Eqs. 5-6); the swing trajectory is then taken as the straight line between these endpoints, and height constraints (Eq. 7) are imposed only along that line. No swing-foot dynamics, kinematics, or continuous collision constraints are part of the optimization. The paper's foot-avoidance claim therefore depends on the Operational Space Controller (Sec. II-F) tracking this straight-line reference within the padded collision radius (0.4 m diameter for a 0.2 m obstacle, Sec. IV.A.4). This assumption is load-bearing: if the actual foot trajectory deviates laterally from the planned line, the height constraint is evaluated at the wrong x-y location, so the foot can clip the obstacle even when the planned line is safe. The paper provides no tracking-error data, no Monte Carlo over obstacle placements or speeds, and only a single qualitative hardware step-over. The authors themselves list \"introducing swing foot kinematics/dynamics via nonlinear MPC\" as future work (Sec. V), effectively conceding that the current formulation omits foot kinematics. Without evidence that tracking error stays within the padding, the central claim of \"foot-level obstacle avoidance\" is not established beyond the specific demonstrated case.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a real-time MPC framework for dynamic bipedal robots that addresses both body-level and foot-level obstacle avoidance. For body avoidance, the paper extends a prior half-space relaxation formulation with an ellipse-based collision region that incorporates obstacle velocity, and it introduces a heuristic that shortens the upcoming stance duration when the MPC's predicted foothold tracking error exceeds a threshold. For foot avoidance, the ground around a rectangular obstacle is decomposed into four convex quadrants, a Mixed-Integer Quadratic Program selects which quadrant the landing foothold must lie in, and step-over height constraints are imposed along the straight line connecting the swing foot's start and end positions. A soft minimum-travel-distance constraint is added to prevent stalling in local minima. The framework is evaluated in multibody simulations on Cassie and Digit, and in a single hardware experiment on Digit where the robot steps over a 0.2 m obstacle.","tokens_in":10693,"tokens_out":3400,"duration_ms":33576,"significance":"If the claims hold, the framework would be a useful contribution to reactive bipedal locomotion: it is, to the authors' knowledge, the first model-based 3D dynamic bipedal walking approach that implicitly chooses to step over or navigate around foot-level obstacles based on CoM dynamics, and the adaptive step-timing heuristic is a plausible mechanism for faster dodging. The formulation is mostly convex with binary variables, the use of an existing Operational Space Controller for tracking is appropriate, and the paper demonstrates the complete pipeline in simulation and on Digit hardware. However, the validation is largely qualitative, and the central foot-avoidance safety claim rests on an unverified assumption about low-level tracking accuracy. The paper would be significantly strengthened by quantitative tracking-error data, repeated trials, and a clearer statement of the conditions under which the foot-avoidance guarantee holds.","major_comments":[{"comment":"The x-y plane foot-avoidance constraints only enforce that the swing foot's start and end footholds lie in the selected convex quadrant; the swing trajectory is then taken as the straight line between these endpoints, and the height constraints of Eq. (7) are imposed only along that line. Consequently, the 3D collision-free guarantee depends on the low-level OSC tracking this straight-line reference with an error smaller than the padded collision region diameter (0.4 m for a 0.2 m obstacle, Sec. IV.A.4). This is a load-bearing assumption, but the paper provides no tracking-error data, no perturbation or Monte Carlo study, and only a single qualitative hardware step-over. Since the authors list 'introducing swing foot kinematics/dynamics via nonlinear MPC' as future work in Sec. V, the current formulation itself omits swing-foot kinematics. I recommend providing quantitative tracking-error measurements during swing (e.g., lateral deviation under the OSC), or a formal bound on the required padding, or clearly softening the central claim to 'demonstrated on the tested configuration.'","section":"Sec. II-E.1, Eqs. (5)-(7)"},{"comment":"The experimental validation is thin relative to the strength of the claims. Cassie results in Fig. 6 are reported only as success or failure, with no quantitative metrics such as minimum distance to the obstacle, CoM tracking error, or the actual step-timing adjustments. The Digit hardware result in Fig. 7B is a single successful step-over, with no reporting of commanded versus actual foot/CoM trajectories, obstacle clearance, or repeated trials. The paper claims that the framework 'implicitly chooses to step over or navigate around' obstacles and that adaptive step timing improves avoidance; without systematic experiments with multiple obstacle sizes, speeds, placements, and a baseline without adaptive timing and without MIQP, these claims cannot be distinguished from a tuned demonstration. I suggest adding repeated trials and quantitative comparison metrics.","section":"Sec. III-C and Sec. IV.A.4"},{"comment":"The adaptive step-time heuristic is derived from the authors' own MPC behavior: the switching threshold is based on the MPC's predicted foothold tracking error, and the decision rule (shorten the next stance phase by 0.1 s when the error exceeds a threshold) is not accompanied by any analysis of when this is guaranteed to improve avoidance or how the threshold was selected. The paper reports no sensitivity analysis for the threshold, minimum stance duration, or discrete step-time interval. This is a design choice rather than a logical error, but it weakens the claim that the framework reduces reliance on human operators, since the heuristic introduces new hand-tuned parameters.","section":"Sec. II-D.2"}],"minor_comments":[{"comment":"In the general MPC formulation, the symbols q, u, rho, and b are used before being defined; please define them in that section or refer explicitly to the later subsection where they are introduced.","section":"Sec. II-B"},{"comment":"The minimum-travel-distance constraint depends on gains K_x and K_y, but the paper does not state how these are chosen or whether they are tuned separately for Cassie and Digit. Please report the values used in the experiments.","section":"Sec. II-E.2, Eq. (11)"},{"comment":"The 'small offset term on the order of 0.01' added to the foot position is not specified with units, and it is unclear whether it is a fixed offset or state-dependent. Please clarify.","section":"Sec. IV.A.2"},{"comment":"The paper states that the MPC is updated at 50 Hz for MIQP and 200 Hz otherwise, but it does not report solver solve times or whether the real-time deadlines were consistently met. Please provide this information.","section":"Sec. IV.A.1"},{"comment":"The related-work discussion mentions [28] as using spline-based swing trajectories, but it does not explicitly state how the present formulation differs with respect to swing-foot optimization. A sentence clarifying that distinction would help position the contribution.","section":"Sec. I-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely publishable after revision, but the main gap is evidence for the foot-avoidance safety guarantee. A possible novelty concern for the editor: the MIQP foothold-selection method is closely related to [28], and the paper would benefit from a more explicit differentiation. The lack of quantitative experimental data may also be a concern for the journal's standards."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read on arXiv:2505.13715.\n\nThe genuinely new piece is the combination: an LIPM-MPC that handles fast-moving obstacles with an ellipse-shaped collision region (using obstacle velocity), selects footholds via MIQP among convex ground quadrants so the robot implicitly steps over or around foot-level obstacles, and adapts step timing based on a predicted foot-tracking error threshold. The adaptive step-time heuristic is the cleverest idea in the paper: it uses the MPC's own predicted tracking error to decide when the next stance phase should be shortened. The anti-stall min-travel-distance constraint is also a practical fix for a real local-minima problem in half-space avoidance. The writing is clear and the authors are upfront about what is simplified.\n\nThe soft spots are real but not fatal. The foot-avoidance guarantee rests on the assumption that the low-level OSC tracks the straight-line swing reference within the padded collision radius (0.4 m for a 0.2 m obstacle). The MPC only constrains endpoints in the x-y plane and height along the planned line; nothing prevents lateral tracking error from pushing the foot into the obstacle. The paper reports no tracking-error data, no repeated trials, and only one qualitative hardware step-over. The authors list swing-foot kinematics via NMPC as future work, which is effectively conceding the gap. That means the central claim is demonstrated for the specific test shown, not established generally. A reviewer should ask for tracking-error plots and a few obstacle placements and speeds.\n\nThe adaptive step-time threshold and several weights are not fully reported, and there is no code or data release. The 'first' claim is plausible but there is no direct benchmark against [28], the closest prior work. None of this is disqualifying for an engineering paper; it's the difference between a solid conference contribution and a journal-level validation.\n\nWho is this for? People working on real-time bipedal MPC, footstep planning, and obstacle avoidance will get value. It deserves a serious referee: the formulation is reproducible in principle, the ideas are well-motivated, and the hardware demo is a genuine data point. I would accept it for peer review and push for a quantitative follow-up.\n\nI would probably not cite it myself in the next year, but I'd read the revised version.","headline":"A well-engineered MPC extension for bipedal body and foot avoidance with a clever adaptive step-time heuristic, but the foot-safety guarantee leans on an unverified tracking margin and the validation is mostly qualitative.","tokens_in":11189,"tokens_out":2582,"would_cite":false,"duration_ms":24144,"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":"This paper presents a real-time MPC framework that lets a bipedal robot dodge fast-moving obstacles and, in the same optimization, choose whether to step over or step around foot-level obstacles.","keywords":["Model Predictive Control","bipedal walking","obstacle avoidance","Mixed-Integer Quadratic Programming","adaptive step timing","foot placement","Linear Inverted Pendulum Model"],"falsifier":"Instrument Digit's swing foot during a step-over of a 0.2 m obstacle with the controller's 0.4 m padded region, and log the minimum vertical and lateral clearance over the obstacle across many trials. If measured clearance ever falls below roughly 0.1 m, the padding on each side, while the MPC still outputs a step-over plan, then the simplified straight-line swing model does not guarantee the advertised 3D avoidance.","tokens_in":10110,"feed_emoji":"🤖","tokens_out":5049,"duration_ms":43960,"temperature":0.7,"pith_summary":"The paper claims that one real-time MPC can handle both body-level and foot-level collision avoidance for dynamic bipedal walking, and that the foot-avoidance layer is the first model-based approach to implicitly choose between stepping over and stepping around a ground obstacle based on the state of the center-of-mass dynamics. The practical payoff is that a walking robot no longer needs a hand-tuned step frequency or a separate footstep planner: the optimizer adjusts step timing, selects a foothold region, and shapes the swing-foot height profile together. The authors back the claim with multibody simulations on Cassie and Digit and with Digit hardware experiments showing a step-over maneuver. A sympathetic reader would take the central contribution to be the integration: lifting non-convex foot-avoidance into a convex MIQP within a LIPM-based MPC while keeping real-time rates.","feed_headline":"Biped MPC picks step-over or step-around on the fly","feed_subtitle":"Real-time controller adjusts step timing and foothold choice, validated on Cassie simulation and Digit hardware.","key_machinery":"The load-bearing object is the LIPM-based MPC with binaries: state variables are pelvis and foot trajectories; costs penalize velocity tracking at step transitions, nominal foot-to-pelvis distance, foot effort, and slack variables for avoidance; constraints include the LIPM dynamics, half-space body-avoidance constraints, and, for feet, a big-M formulation that selects exactly one of four convex ground quadrants around each obstacle. A soft minimum-travel-distance constraint, encoded with auxiliary positive/negative parts and binaries, forces the prediction to keep moving so the optimizer cannot stall behind an obstacle. Step timing is handled not as a continuous variable but as discrete choices at 0.1 s intervals, switched by a heuristic based on predicted foothold tracking error.","core_discovery":"On its own terms, the paper's discovery is that a Linear Inverted Pendulum Model (LIPM) based MPC can be extended to plan 3D swing-foot avoidance by decomposing the ground around each obstacle into four convex quadrants, selecting one quadrant with binary variables in a Mixed-Integer Quadratic Program, and imposing height constraints along the straight line connecting the current and selected foothold. The same MPC also adjusts step timing: a tracking-error heuristic shortens the next stance phase when the foothold error predicts that staying on that foot would block a dodge, and the collision region is modeled as a velocity-stretched ellipse so the avoidance direction accounts for where the obstacle is going. The authors state that, to their knowledge, this is the first documented model-based approach for 3D dynamic bipedal walking that implicitly chooses to step over or navigate around foot-level obstacles depending on the state of its CoM dynamics.","pith_inferences":["If OSC tracking error stays within the padded margin, the straight-line swing reference makes the 3D clearance guarantee largely a matter of respecting height constraints; a natural extension is replacing the straight line with a spline or adding swing-foot kinematics directly into the MPC, which the paper lists as future work.","The velocity-stretched ellipse could be reused for moving obstacles with uncertain trajectories by inflating the ellipse with the obstacle's speed and heading, though the paper only demonstrates it with a known linear motion.","The tracking-error heuristic for step timing might transfer to other bipedal tasks that need fast weight transfer, such as push recovery, if the same error signal can be computed there.","The convex-quadrant decomposition could be generalized to multiple ground obstacles with a more advanced decomposition method, as the paper notes, which would extend the same MIQP mechanism to cluttered terrains."],"forward_implications":["A robot using this controller can dodge a 5 m/s obstacle by shortening its next stance phase to 0.2 s instead of 0.4 s, as shown in Cassie simulation.","The same MPC will step around at low commanded speed and step over at higher speed, because the MIQP region choice trades travel distance against avoidance cost.","Foot-level obstacle negotiation runs in real time: MPC at 50 Hz with MIQP and 200 Hz otherwise, on Digit hardware.","The controller requires no a priori switch between 'step over' and 'step around' behaviors; the behavior emerges from the optimization.","The three-dimensional avoidance guarantee is tied to the straight-line swing reference, so the method's clearance can be assessed directly from the padded obstacle region."],"supporting_citations":[{"why":"Supplies the baseline LIPM-based MPC for body avoidance that this paper extends with adaptive step timing and foot avoidance.","marker":"[18]"},{"why":"Provides the Linear Inverted Pendulum Model used as the reduced-order prediction dynamics.","marker":"[12]"},{"why":"Prior MIQP foothold-selection formulation whose region decomposition and big-M constraints are adapted here.","marker":"[28]"},{"why":"Operational space controller used to track the MPC reference trajectory on the robot.","marker":"[31]"},{"why":"Gurobi solver used for the mixed-integer problems, enabling real-time MIQP solutions.","marker":"[32]"},{"why":"MuJoCo physics engine used for the Digit multibody simulations.","marker":"[33]"},{"why":"Contact-aided Kalman filter used for state estimation in the controller loop.","marker":"[34]"},{"why":"OSQP solver used for the non-integer quadratic programs in the framework.","marker":"[35]"},{"why":"CasADi used to generate the C++ implementation of the MPC.","marker":"[36]"}],"fun_headline_variants":["Biped MPC dodges obstacles with MILP foothold picks","Real-time biped MPC steps over or around on the fly","MILP-based biped MPC chooses step-over or step-around","Adjusts step timing and foothold in real time","LIPM MPC with MILP picks step-over or step-around"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 3D foot-avoidance guarantee relies on the lower-level controller tracking the planned straight-line swing path closely enough that the padded collision region absorbs all tracking error, since the MPC itself never models the swing foot's dynamics or kinematics.","fun_headline_variants_meta":{"raw":{"variants":["Biped MPC dodges obstacles with MILP foothold picks","Real-time biped MPC steps over or around on the fly","MILP-based biped MPC chooses step-over or step-around","Adjusts step timing and foothold in real time","LIPM MPC with MILP picks step-over or step-around"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000826,"raw_usage":{"total_tokens":3612,"prompt_tokens":951,"completion_tokens":2661,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2575}},"tokens_in":567,"tokens_out":2661,"duration_ms":17923,"temperature":1.0,"reasoning_tokens":2575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:11:13.547329+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument Digit's swing foot during a step-over of a 0.2 m obstacle with the controller's 0.4 m padded region, and log the minimum vertical and lateral clearance over the obstacle across many trials. If measured clearance ever falls below roughly 0.1 m, the padding on each side, while the MPC still outputs a step-over plan, then the simplified straight-line swing model does not guarantee the advertised 3D avoidance.","supporting_citations":[{"cited_title":"Real-time dynamic bipedal avoidance,","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline LIPM-based MPC for body avoidance that this paper extends with adaptive step timing and foot avoidance."},{"cited_title":"The 3d linear inverted pendulum mode: a simple modeling for a biped walking pattern generation,","cited_arxiv_id":null,"evidence_quote":"Provides the Linear Inverted Pendulum Model used as the reduced-order prediction dynamics."},{"cited_title":"Bipedal walking on constrained footholds with mpc footstep control,","cited_arxiv_id":null,"evidence_quote":"Prior MIQP foothold-selection formulation whose region decomposition and big-M constraints are adapted here."},{"cited_title":"Fast online trajectory optimization for the bipedal robot cassie,","cited_arxiv_id":null,"evidence_quote":"Operational space controller used to track the MPC reference trajectory on the robot."}],"review_version":1}