{"id":"7c26a520-7a43-41bd-a6e8-9f2f3e42cd24","arxiv_id":"2506.16986","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A legged mobile manipulator throws grasped objects to 4-6 m targets with mean landing errors around 0.28-0.43 m, using a learned nominal policy, a 400 Hz residual policy, and closed-loop pullback tube acceleration.","lead":"This paper reports a control system that lets a four-legged robot with an arm grip and throw objects such as snowballs and floorballs, hitting targets several meters away with roughly 30 cm accuracy. It combines a learned whole-body tracking policy with a fast corrective network and a math-based acceleration smoother for release-time uncertainty.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pullback tube's release-window kinematics in Problem Tube-CVX omit the 0.5*T^2*a_tube position term, producing a biased landing prediction that the claimed tube contribution inherits.","rationale":"The reader's weakest assumption (drag and detach-time bounds) is empirical and reasonable, but it does not address the internal consistency of the tube acceleration module that carries the paper's robustness claim. My focus is the central claim that the full framework, including the pullback tube, produced the reported accuracy. The hardware numbers are plausible and the authors disclose partial checkpoints and sim-to-real gap, which supports a conditional verdict. However, the kinematic model in Problem Tube-CVX is internally inconsistent as written: the position update ignores the acceleration term that the velocity update includes over the same horizon. This is not a modeling simplification; it is a mismatch in the core release-window dynamics. If the corrected term changes the optimizer's decisions, then Sec. IV.D and the tube's contribution to Sec. IV.E are not reliable. The proposed re-simulation is a concrete, bounded check that settles the issue. I therefore keep the reader's conditional verdict rather than moving to rejection, because the empirical results could still hold even if the written formulation needs correction.","tokens_in":11477,"tokens_out":12523,"duration_ms":135556,"concrete_test":"Re-derive Problem Tube-CVX with p_T = p_EE + T v_EE + 0.5 * T^2 * a_tube, keeping all other constraints and the same Sec. IV.D perturbation mesh (Table III, 1500 initial states, 5 seeds). Re-run the simulation and compare the optimized a_tube commands and the maximum landing error in the 50-100 ms detach window against Table IV. If the corrected predictor shifts the maximum landing error by more than roughly 5 cm or materializes a different acceleration pattern, the published tube module is not solving the stated problem and the Sec. IV.D/E results require re-evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. III.C.2, Problem Tube-CVX predicts the terminal EE state under a commanded constant tube acceleration a_tube using p_T = p_EE + T v_EE and v_T = v_EE + T a_tube. If a_tube is a constant acceleration applied over the remaining release window T, the exact terminal position is p_T = p_EE + T v_EE + 0.5 * T^2 * a_tube. The formulation omits this second-order term, then linearizes the landing map only with respect to the velocity change. The omitted term is not negligible at the stated operating conditions: for T = 0.1 s and a_tube = 10 m/s^2, the predicted release position shifts by 0.05 m, and by 0.10 m at 20 m/s^2, which is on the order of the claimed 0.276 m landing error. Sec. IV.D, Table IV, and the 6.17% improvement in Sec. IV.E all inherit this prediction bias unless p_T is intended as a constant-velocity reference rather than the true resulting position, a distinction the paper does not state. If the bias is real, the convex program solves a different problem from the actual robust-throwing problem, and the tube module's contribution is not what is being evaluated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-component control framework for whole-body prehensile throwing with a legged mobile manipulator: a 100 Hz nominal tracking policy, a 400 Hz residual policy, and a real-time convex \"pullback tube acceleration\" optimizer based on a backward reachable tube. The authors report hardware landing errors of 0.276 m at 6 m and 0.429 m at 4 m over 40 throws, a simulated ablation showing the full method improves landing accuracy by 20.04% over the nominal policy, and a human-comparison experiment in which the robot achieved 71/125 successes versus 19/125 for 25 students. The abstract also claims a 0.398 m/s velocity tracking error and a 56.8% success rate. The paper's central assertion is that this is the first reported whole-body prehensile throwing with quantified accuracy on hardware.","tokens_in":11617,"tokens_out":11736,"duration_ms":109818,"significance":"If the reported results hold, this is a useful early demonstration of whole-body dynamic throwing on a legged manipulator, combining learning-based tracking with model-based robustification. The 400 Hz residual policy and the closed-loop, 0.4 ms convex tube-acceleration module are technically interesting, and the hardware experiments with diverse objects (gift box, snowball, floorball) go beyond fixed-base throwing studies. The paper also provides quantified accuracy metrics, which prior legged throwing work lacked. However, the significance is currently bounded by several experimental and modeling gaps detailed below, especially the self-acknowledged use of partial implementations for the human comparison and the inconsistency in the tube-acceleration kinematics.","major_comments":[{"comment":"Problem Tube-CVX has two related technical issues. First, the terminal position constraint p_T = p_EE + T v_EE omits the 0.5*T^2*a_tube term: for a constant tube acceleration a_tube applied over the remaining release window T, the exact terminal position is p_T = p_EE + T v_EE + 0.5*T^2*a_tube. At T = 0.1 s and a_tube = 10 m/s^2 the omitted term is 0.05 m, and it reaches 0.10 m at 20 m/s^2, i.e., on the order of the claimed 6 m landing error of 0.276 m. Because the landing map is then linearized only with respect to the velocity change, the optimizer solves a biased problem unless p_T is explicitly defined as a constant-velocity placeholder rather than the true predicted release position. Second, the equality constraint ˙r_T = ||v_T,xy||_2 is nonlinear and nonconvex, contradicting the statement that all equality constraints are linear; as written, the program is not convex, so the claimed 0.4 ms solve time and the convexity-based guarantees need re-examination. Please correct the formulation or justify both approximations quantitatively.","section":"III.C.2"},{"comment":"The human-comparison result is compromised by the note immediately before Section IV.A, which states that \"the comparison of accuracy with human participants were performed under such conditions,\" i.e., with partial implementations of the proposed method or earlier training checkpoints. The abstract nevertheless presents the 56.8% versus 15.2% success rates as a system-level comparison. Since this comparison may not use the full proposed framework, it cannot serve as evidence for the complete method. Please specify exactly which implementation and checkpoint were used, or re-run the comparison with the final system.","section":"IV.A"},{"comment":"The hardware accuracy results are reported only as mean landing errors over 40 throws (10 per target location at 4 m and 6 m with lateral offsets ±0.5 m). No per-target means, standard deviations, confidence intervals, or raw data are provided. Without these, the 0.276 m and 0.429 m figures and the claimed 49.5% improvement over the nominal policy cannot be statistically evaluated. Please add per-target statistics and variance information.","section":"IV.B"},{"comment":"The ablation does not isolate the contribution of the pullback tube acceleration module. The description states that variations without the tube optimizer were \"instructed to track this constant velocity,\" while the full method tracked the varying EE velocities generated by the tube optimizer. The 6.17% landing-error improvement attributed to the tube module therefore conflates the effect of a different reference trajectory with the effect of the closed-loop tube acceleration. A cleaner ablation would use the same optimized reference commands without the closed-loop tube correction, or otherwise control for the reference change.","section":"IV.E"},{"comment":"The \"Residual 100 Hz\" row reports a success rate of 13.9%, which is inconsistent with its other metrics: it has lower position tracking error (0.0466 vs 0.0693), lower velocity tracking error (0.2269 vs 0.2506), lower L1 landing error (0.2139 vs 0.2530), and lower landing MSE (0.0692 vs 0.1658) than the nominal policy, yet the success rate drops from 63.3% to 13.9%. Unless the error distributions are pathologically heavy-tailed, this appears to be a typographical error (perhaps 63.9%). Please correct or explain this discrepancy, as the table is central to the ablation claim.","section":"Table V"}],"minor_comments":[{"comment":"The abstract states a \"velocity tracking error of 0.398 m/s\" but this number does not appear in Section IV.F or elsewhere in the results; please report this metric with its conditions (commanded velocity, number of throws, variance) or remove it from the abstract.","section":"Abstract/IV.A"},{"comment":"The abstract says the human comparison used targets \"randomly placed at distances of 3-5 m,\" while Section IV.A says the target was placed \"3-4 meters away\"; please make the distance range consistent.","section":"IV.A"},{"comment":"The procedure for measuring landing positions on hardware is not described; please state the measurement method and its accuracy.","section":"IV.B"},{"comment":"The batch simulation in Section IV.D evaluates the pullback tube using the same drag-free projectile model that was used to construct the BRT and nominal throwing velocities, so it does not assess robustness to unmodeled aerodynamic drag; a sensitivity study with a different flight model (e.g., quadratic drag) would strengthen the robustness claim.","section":"IV.D"},{"comment":"The figure caption reads \"The proposed control framework.consists\" with a missing space after the period.","section":"Fig. 2"},{"comment":"The conclusion's limitation statement that the residual policy only tracks vertical acceleration and overfits the simulation should be reflected more explicitly in the abstract's general accuracy claims.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The self-acknowledged partial-implementation note for the human comparison is unusual and should be treated as an explicit limitation; the authors should be asked to state clearly which experiments use the complete method. The central technical concern is the tube-CVX kinematics, which appears to be a genuine modeling error rather than a presentation issue, and the nonconvex equality constraint further undermines the convexity claim. The inconsistent success rate in Table V should also be resolved before the paper is reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the hardware demonstration: whole-body prehensile throwing on a legged manipulator with reported landing errors of 0.276 m at 6 m and 0.429 m at 4 m over 40 throws. That is worth knowing, and the integration of a 400 Hz residual policy with an online-optimized tube acceleration is a sensible engineering move. The authors are honest about several limitations: partial checkpoints, drag neglect, and the negligible-object-mass assumption are all disclosed.\n\nThe soft spots are real, though. The most serious is the kinematics in Problem Tube-CVX. The paper defines p_T = p_EE + T v_EE and v_T = v_EE + T a_tube. If a_tube is a constant acceleration applied over the remaining release window, the exact terminal position is p_T = p_EE + T v_EE + 0.5 T^2 a_tube. The omitted term is not small: at T = 0.1 s and a_tube = 10 m/s^2 it is 0.05 m, and at 20 m/s^2 it is 0.10 m, both on the order of the claimed landing error. The following linearization only corrects for the velocity change, so the optimizer is solving a biased problem. That undermines the attribution of the 6.17% improvement in Sec. IV.E and the simulation results in Table IV. The paper asserts this convex program is a tight relaxation of the robust throwing problem, but as written it is not the formulation from Liu and Billard [5].\n\nOther issues are less severe but worth naming. The central claim that the pullback tube makes the BRT an attracting invariant set is asserted without proof. The hardware accuracy numbers come from 40 throws with no error bars or per-target variance. The abstract reports a velocity tracking error of 0.398 m/s that I could not find reported in the results section. The human comparison was run under a partial implementation and an informal protocol, so the 71/125 vs 19/125 result is anecdote, not benchmark.\n\nFor a reader tracking legged manipulation or dynamic throwing, this paper is worth a look as an early quantified demonstration. But the theory needs fixing before the tube-acceleration claims can be trusted. I would send it to peer review—a good referee will catch the kinematic error and push for a corrected formulation with error bars. As it stands, I would not cite the tube module's results in my own work.","headline":"Solid hardware demonstration of whole-body throwing with a legged robot, but the tube-acceleration optimizer contains a kinematic omission that biases the central results.","tokens_in":12309,"tokens_out":4350,"would_cite":false,"duration_ms":41075,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A legged robot throws with 0.28 m accuracy at six meters","keywords":["whole-body manipulation","prehensile throwing","legged mobile manipulators","reinforcement learning","residual policy","backward reachable tube","tube acceleration","release uncertainty"],"falsifier":"Record the actual detach instant with a high-speed camera and the landing positions with motion capture for 40 throws at six meters using a light foam ball; if mean landing error exceeds about 0.5 meters or detach times fall outside 50-100 milliseconds, the drag-free and release-window assumptions are violated in that regime.","tokens_in":11130,"feed_emoji":"🎯","tokens_out":8165,"duration_ms":75101,"temperature":0.7,"pith_summary":"The paper argues that accurate whole-body throwing on a legged mobile manipulator can be achieved by combining a learned end-effector tracking policy with a model-based robust release controller. Its central evidence is hardware: an average landing error of 0.28 meters when throwing grasped objects at targets six meters away, and a 56.8 percent success rate against 15.2 percent for human students when hitting small targets at 3-4 meters. A reader should care because dynamic throwing lets mobile robots deliver or place objects beyond arm reach, and the work is presented as the first quantified whole-body prehensile throwing demonstration on hardware. The same evidence also shows the claim is bounded: without the residual policy and pullback acceleration module, landing error roughly doubles.","feed_headline":"A legged robot throws with 0.28 m accuracy at six meters","feed_subtitle":"Learned tracking plus release-motion control lets it hit small targets more often than human students.","key_machinery":"The load-bearing object is the backward reachable tube (BRT), the set of release states whose projectile trajectories can reach the landing target, together with the pullback tube acceleration optimizer. At 400 Hz, the optimizer solves a convex program that, given the current end-effector state and the time left in the release window, commands a constant acceleration tube that steers the state into and keeps it inside the BRT, making the landing position largely insensitive to when the gripper actually releases the object. Around this sit a 100 Hz nominal policy that tracks a commanded throwing trajectory and a 400 Hz residual policy that refines the joint targets; the residual policy matters most at high throwing velocities, and the convex solver is fast enough (about 0.4 milliseconds) to run in closed loop.","core_discovery":"The central claim is that prehensile whole-body throwing can be reformulated as a high-accuracy end-effector velocity tracking problem, and that a 400 Hz residual policy plus an optimization-based pullback tube acceleration module makes that tracking accurate enough to land objects within about 0.28 meters at six meters. The full system reports mean landing errors of 0.276 meters at six meters and 0.429 meters at four meters over 40 hardware throws, a 49.5 percent error reduction over the nominal policy alone, and a 71/125 versus 19/125 score in a corridor test against 25 students. The authors explicitly claim this is the first reported whole-body prehensile throwing with quantified accuracy on hardware.","pith_inferences":["Editorial inference: because the tube's backward reachable set is computed from drag-free projectile motion, lightweight and deformable objects are the likely failure boundary; replacing the flow map with a drag-aware or learned model is a natural next step.","Editorial inference: the pullback tube optimizer is a generic release-layer controller that could be attached to any policy that can report end-effector state at 400 Hz, not only the one trained in this paper.","Editorial inference: the human comparison should not be read as evidence that humans throw poorly; untrained students with no practice are a weak baseline, and a controlled comparison with practiced throwers would be the informative benchmark.","Editorial inference: the larger mean error at four meters (0.429 m) than at six meters (0.276 m) hints at a systematic bias in nominal velocity or release timing rather than pure noise, which a high-speed camera measurement of detach time could isolate."],"forward_implications":["Whole-body throwing on legged robots can be treated as a timed end-effector velocity tracking problem, with robust release-motion synthesis absorbing release-timing uncertainty in the 50-100 millisecond window.","At high release speeds (7-10 m/s), the residual policy reduces velocity tracking error by up to 16.8 percent over the nominal policy, so high-frequency refinement mainly pays off in fast throws.","The legged base measurably contributes to throwing power, delivering about 53.4 percent more angular impulse than a tabletop arm executing the same joint trajectory.","Simulation ablations put the full method's landing-error improvement at 20.04 percent over the nominal policy and 6.17 percent over the residual-policy-only variant for 7 m throws."],"supporting_citations":[{"why":"Supplies the residual-physics throwing baseline and establishes the accuracy bar for data-driven throwing on fixed arms.","marker":"[1]"},{"why":"Documents prior whole-body throwing on legged robots without quantified hardware accuracy, the gap this paper addresses.","marker":"[3]"},{"why":"Provides the backward-reachable-tube and tube-acceleration theory for robust dexterous throwing that the optimization module adapts to closed-loop use.","marker":"[5]"},{"why":"Introduces residual policy learning, the technique the high-frequency residual policy extends.","marker":"[15]"},{"why":"Supplies the multi-throw training episode structure used to train the nominal policy.","marker":"[18]"},{"why":"Reports that a directly deployed RL policy robustified only by randomization produced dangerous motion, motivating explicit robust release-motion synthesis.","marker":"[19]"},{"why":"Explains why analytical flight solutions generally fail with nonlinear drag, supporting the paper's drag-free simplification for the BRT.","marker":"[21]"},{"why":"Ensures the tube-acceleration program can be expressed in disciplined parametrized programming, enabling embedded code generation.","marker":"[24]"},{"why":"Enables the roughly 0.4 ms solver time that lets the pullback tube run in closed loop at 400 Hz.","marker":"[25]"}],"fun_headline_variants":["Legged robot throws 6 m with 0.28 m accuracy, beats humans","Throwing robot hits small targets 56.8% of time; humans 15.2%","0.28 m landing error at 6 m: learned whole-body throwing","Whole-body throwing accuracy: 0.28 m at 6 m, better than students","Residual policy + tube acceleration cut throw error by 49.5%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The release plan assumes the object flies like a drag-free projectile and that the gripper releases the object 50-100 milliseconds after the opening command; both can fail for light, deformable, or slippery objects and for worn grippers.","fun_headline_variants_meta":{"raw":{"variants":["Legged robot throws 6 m with 0.28 m accuracy, beats humans","Throwing robot hits small targets 56.8% of time; humans 15.2%","0.28 m landing error at 6 m: learned whole-body throwing","Whole-body throwing accuracy: 0.28 m at 6 m, better than students","Residual policy + tube acceleration cut throw error by 49.5%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000355,"raw_usage":{"total_tokens":1894,"prompt_tokens":877,"completion_tokens":1017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":906}},"tokens_in":493,"tokens_out":1017,"duration_ms":8910,"temperature":1.0,"reasoning_tokens":906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:15:11.815266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the actual detach instant with a high-speed camera and the landing positions with motion capture for 40 throws at six meters using a light foam ball; if mean landing error exceeds about 0.5 meters or detach times fall outside 50-100 milliseconds, the drag-free and release-window assumptions are violated in that regime.","supporting_citations":[{"cited_title":"A review of the analysis of wind- influenced projectile motion in the presence of linear and nonlinear drag force,","cited_arxiv_id":null,"evidence_quote":"Explains why analytical flight solutions generally fail with nonlinear drag, supporting the paper's drag-free simplification for the BRT."},{"cited_title":"Differentiable convex optimization layers,","cited_arxiv_id":null,"evidence_quote":"Ensures the tube-acceleration program can be expressed in disciplined parametrized programming, enabling embedded code generation."}],"review_version":1}