{"id":"8645d26c-18f9-4a11-a23d-f7e9c04d50da","arxiv_id":"2504.14989","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hierarchical reinforcement learning controller switches between dribbling and locomotion skills to move a ball across rugged terrain.","lead":"This paper describes a two-level controller that lets a four-legged robot dribble a soccer ball across ramps, stairs, and bumpy ground. A high-level policy switches between separately trained walking and dribbling skills; the authors report improved success over earlier dribbling systems, but the evaluation has important gaps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The simulation success metric in Table II is defined as the robot reaching the terrain boundary (Sec. VI.A.2), so it does not require transporting or controlling the ball; the central outperformance claim is therefore unsupported by the reported measurements.","rationale":"The reader identified the simulation success metric as the weakest assumption, and I agree. The central claim is not just that the robot can walk over rugged terrain, but that it outperforms prior work in dynamic ball manipulation. Table II is the principal simulation evidence for this claim, and its stated success criterion ignores the ball entirely. This makes the headline comparison uninterpretable as a manipulation result. The DSF-PO derivation also lacks formal rigor, but even if the loss were theoretically sound, the reported evaluation would still not establish the central contribution. The real-world experiments are too small and use externally reported baseline numbers, so they cannot independently carry the claim. Since the reader's rejection is based on this same load-bearing measurement problem, I do not see a reason to change the verdict; the appropriate outcome remains rejection unless the authors re-evaluate with a ball-aware metric and fair, directly rerun baselines.","tokens_in":10880,"tokens_out":4674,"duration_ms":45378,"concrete_test":"Re-run the Table II experiment with the same checkpoints, command, terrains, and 100 trials per condition, but replace the success criterion with a ball-aware one: success iff the robot crosses the terrain boundary and the ball also crosses the same boundary within a defined lateral corridor (e.g., 2 m) and within 3 m of the robot at that moment. Additionally report the per-trial fraction of episode time during which the ball is within 0.5 m of the robot. If Ours no longer exceeds both baselines on all four non-flat terrains, the outperformance claim fails. If code or checkpoints are unavailable, the authors should at least provide per-trial logs of ball displacement and ball-velocity tracking error for the original experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VI.A.2 states that success is defined as the robot reaching the terrain boundary, and Table II labels these results as ball dribbling success rates. Under this metric, a policy can score a success by abandoning the ball and walking to the boundary, because no condition requires the ball to cross the boundary, remain near the robot, or be controlled at all. The comparison against DribbleBot and DexDribbler therefore measures terrain traversability of the locomotion components, not dynamic ball manipulation on rugged terrain. This is the load-bearing measurement for the paper's central claim that the method outperforms baseline approaches in dynamic ball manipulation. The cross-terrain demo in Fig. 4 is a single uncompared trajectory, and the real-world comparison in Table III uses published baseline numbers with only five trials per terrain. Neither repairs the core problem: the main quantitative superiority claim can be satisfied even if the robot fails the manipulation task. The DSF-PO issue (Eq. 7 is admitted to be only 'partially correct' in intuition) is relevant to the novelty claim, but the metric problem alone is sufficient to invalidate the central empirical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a hierarchical reinforcement learning (RL) framework for dynamic legged ball manipulation on rugged terrain. A high-level policy, trained with a proposed Dynamic Skill-Focused Policy Optimization (DSF-PO), selects among four pre-trained low-level skills (two dribbling, two locomotion) and issues skill-specific commands. The framework is trained in Isaac Gym and transferred zero-shot to a Unitree Go2. The central empirical claim is that the proposed method outperforms DribbleBot and DexDribbler in ball manipulation on rugged terrain, based on simulation results in Table II and real-world results in Table III.","tokens_in":11072,"tokens_out":7632,"duration_ms":70504,"significance":"If the central claim were established, the work would be a useful contribution to learning-based loco-manipulation: it addresses a mixed discrete-continuous hierarchical action space, provides a skill-switching mechanism, and demonstrates real-world deployment. The paper also includes a clear low-level skill decomposition and a plausible curriculum design, and the authors are transparent that baseline real-world reproduction was infeasible. However, the quantitative evidence for the headline outperformance claim is not adequate: the main simulation metric does not actually measure ball manipulation, and the real-world comparison is not controlled. These issues are load-bearing because the abstract and conclusion explicitly claim superiority over baselines in dynamic ball manipulation.","major_comments":[{"comment":"The simulation success metric is defined as \"the robot reaching the terrain boundary,\" with no requirement on the ball's final position, trajectory, or continued contact. A policy that abandons the ball and walks to the boundary would be counted as successful, so Table II, titled \"ball dribbling success rates,\" does not measure ball manipulation. This metric is the basis for the central claim that the proposed method outperforms DribbleBot and DexDribbler in dynamic ball manipulation across rugged terrains. The evaluation should use a ball-centric success criterion, such as the ball crossing the boundary while remaining within a distance threshold of the robot, and this metric should be reported consistently for all compared methods.","section":"Section VI.A.2, Table II"},{"comment":"The real-world comparison is not controlled. The baseline numbers are taken from published results [5], [6] rather than measured in the same environment, on the same robot hardware, or under the same task definition; the text explicitly states that reproducing them is infeasible. With only five trials per condition and no statistical analysis, Table III cannot substantiate the claim that the proposed method achieves the highest success rate. The authors should either reproduce baselines under matched conditions or report absolute success counts without comparative claims.","section":"Section VI.B, Table III"},{"comment":"The derivation of DSF-PO is not self-consistent. According to the policy decomposition in Eq. (4), the likelihood of the sampled command for skill d is N(c_d; mu_d, Sigma_d), so the standard importance ratio for that command should be N(c_d; mu_d)/N_old(c_d; mu_d_old). Eq. (5)/(10) instead raises this ratio to the power w_k(s), and Eq. (7) is admitted in the text to be only \"partially correct\" as an intuitive understanding. Without a corrected derivation from a well-defined objective, the claimed property of suppressing gradients from inactive skills is not established. The Fig. 3 ablation, which only reports total reward and episode length, does not by itself validate the loss formulation.","section":"Section IV, Eqs. (5)-(7)"}],"minor_comments":[{"comment":"The cross-terrain demonstration is a single trajectory with no repeated trials, no baseline comparison, and no explicitly defined success criterion; the text states that the robot \"successfully completes the task\" but does not specify how completion is measured.","section":"Section VI.A.3, Fig. 4"},{"comment":"The shaded regions are described as standard deviations over multiple runs, but the number of random seeds is not stated; this should be specified for reproducibility.","section":"Fig. 3"},{"comment":"The notation C^k and I(c_t^k in C^k) is introduced in Eq. (10) but not defined before first use; clarify how the shared command subsets for dribbling skills are handled.","section":"Eq. (10)"},{"comment":"The description of the five-dimensional command output as a concatenation for dribbling and locomotion skills is not fully aligned with the per-skill distribution formulation in Section IV; the relationship between the shared command subsets and the per-skill normal distributions should be made precise.","section":"Section III.B.2 and Section IV"},{"comment":"The \"-\" entries for stair descent indicate that the baseline methods were not evaluated on that terrain, so the statement that the proposed method achieves the highest success rate \"across all four terrains\" is misleading; this limitation should be stated explicitly.","section":"Table III"}],"recommendation":"reject","confidential_remarks":"The evaluation-metric problem is severe enough that the headline outperformance claim cannot be accepted on the current evidence. The simulation comparison in Table II measures terrain traversability, not ball manipulation, and the real-world comparison in Table III relies on published numbers from uncontrolled conditions. Even if the simulation evaluation were corrected, the real-world outperform claim would require new baseline experiments that the authors state are infeasible. In my view this is a reject rather than a major revision, because the central quantitative contribution is not supported by the reported measurements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The task is genuinely new: getting a quadruped to dribble a ball over ramps, rough ground, and stair descents, with a high-level policy switching between pre-trained dribbling and locomotion skills. The authors built the whole pipeline, trained it in Isaac Gym, and transferred it zero-shot to a real Go2, including a downhill stair case where the robot kicks the ball ahead and follows. That is real work, and the video evidence reportedly supports the qualitative claim. Credit where due: the skill-switching behavior (Fig. 4b) and the terrain-specific usage statistics (Fig. 5) show the hierarchy is doing something sensible, not just oscillating randomly.\n\nThe problem is the measurement. Table II defines success as 'the robot reaching the terrain boundary.' That does not require the ball to go anywhere or stay in contact. Under that metric, a policy that walks to the edge and abandons the ball scores a success. The paper calls these 'ball dribbling success rates,' but they are really terrain traversal rates with a ball present. The baselines—DribbleBot and DexDribbler—were locomotion-and-dribbling policies, so the comparison partly measures whether each approach can traverse terrain, not whether it controls the ball. The cross-terrain demo in Fig. 4 is a single uncompared trajectory, so it does not fix this. The real-world numbers in Table III are better because success there is defined as moving the ball to the other end, but those use published baseline results from different papers under different conditions, with only five trials per terrain. That is a legitimate soft spot, not fatal.\n\nDSF-PO is the weakest intellectual piece. The paper itself admits Eq. (7) is only 'partially correct' as an intuitive gradient statement, and the actual loss in Eq. (10) is a PPO variant with per-dimension weights. It may be a useful engineering trick, but calling it 'Dynamic Skill-Focused Policy Optimization' and presenting it as a contribution oversells it. The ablation in Fig. 3 shows it helps in training, so it is not worthless.\n\nThe central claim—that the method outperforms DribbleBot and DexDribbler in dynamic ball manipulation on rugged terrain—is not supported by the reported numbers. The simulation metric cannot distinguish ball control from abandonment. The real-world comparison is too weak to carry the load. The paper deserves a serious referee because the task, the architecture, and the real-world demonstration are all worth scrutiny, but as submitted the quantitative evidence is not there. I would ask for a re-run with a ball-aware success metric (e.g., ball crossing the boundary, or ball position error below a threshold) and a reasonable baseline comparison in simulation, plus more real-world trials. The authors may well be right that their method works; they just have not shown it yet.","headline":"A useful hierarchical RL system for rugged-terrain dribbling, but the main quantitative claim is undercut by a success metric that doesn't require ball control.","tokens_in":11642,"tokens_out":689,"would_cite":false,"duration_ms":7572,"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 hierarchical RL policy that switches between dribbling and locomotion skills lets a quadruped dribble a ball across rugged terrain, beating single-policy baselines.","keywords":["hierarchical reinforcement learning","quadruped loco-manipulation","ball dribbling","rough terrain navigation","skill switching","mixed discrete-continuous action spaces","sim-to-real transfer","dynamic skill-focused policy optimization"],"falsifier":"Re-run the terrain traversability evaluation with success redefined as the ball crossing the same boundary point as the robot, and count episodes in which the ball is abandoned; if the reported advantage over the baselines disappears under this ball-focused criterion, the central claim of superior ball manipulation is not supported.","tokens_in":1874,"feed_emoji":"⚽","tokens_out":2802,"duration_ms":92952,"temperature":0.7,"pith_summary":"The paper is trying to establish that dynamic ball manipulation on rugged terrain is achievable by a quadruped robot through hierarchical reinforcement learning rather than end-to-end single-policy training. The claim is that a high-level policy, running at 10 Hz and reading the robot's own sensors plus the ball's position, can switch among pre-trained low-level skills — two dribbling skills and two locomotion skills — to keep the ball moving while traversing ramps, rough ground, stairs, and flat terrain. To make this work, the paper introduces Dynamic Skill-Focused Policy Optimization (DSF-PO), a modification of the PPO surrogate loss that weights gradients by the probability of selecting the active skill and masks out command dimensions that are irrelevant to that skill. Simulation and real-world experiments are presented as evidence that this framework outperforms existing dribbling methods on rugged terrain. A sympathetic reader would care because legged loco-manipulation is currently mostly confined to flat ground, and the proposed decomposition is a route to extending it to realistic outdoor environments.","feed_headline":"Quadruped dribbles a ball across ramps, stairs, and gravel","feed_subtitle":"A high-level policy switches between dribbling and walking skills, beating prior methods on rugged terrain.","key_machinery":"The central machinery is the DSF-PO surrogate loss, a PPO objective adapted to mixed discrete-continuous action spaces. In standard PPO the importance ratio is $\\pi_\\theta(a_t|s_t)/\\pi_{\\theta_{\\mathrm{old}}}(a_t|s_t)$; DSF-PO rewrites this as a skill ratio multiplied by command ratios, with each command ratio raised to the skill-focus weight $w_k(s_t)=\\pi^d_\\theta(k|s_t)$ and gated by an indicator so only the command block belonging to the selected skill enters the gradient. The effect is that the optimizer pushes the policy toward higher-probability skills and updates only the command dimensions that those skills actually consume. The second load-bearing piece is the hierarchical execution loop: the high-level actor runs at 10 Hz, the low-level skills run at 50 Hz producing joint targets for PD control, and during high-level training the low-level skills and the context-aided estimator are frozen.","core_discovery":"On its own terms, the paper establishes that ball dribbling and terrain traversal need not be learned as one monolithic policy. The contribution is a two-level architecture: four low-level skills are trained end-to-end in simulation — two dribbling skills that apply kicks of different amplitudes and two locomotion skills for fast flat movement and rugged-terrain traversal — and a high-level actor outputs a discrete skill index and continuous low-level commands at 10 Hz. The high-level policy is trained with DSF-PO, whose importance ratio separates the discrete skill-selection ratio from the continuous command ratio and weights each command block by the skill focus weight, so that inactive command dimensions (for example locomotion commands while dribbling) do not produce gradient noise. The paper reports that the resulting policy reaches higher cumulative reward and longer episodes than standard PPO in training, reaches the terrain boundary more often than two prior dribbling systems on ramp-up, ramp-down, rough, and stair-descent terrain, and transfers zero-shot to a physical quadruped that dribbles a ball across indoor ramps, stairs, and outdoor gravel.","pith_inferences":["A stricter test of the paper's central claim would track whether the ball crosses the terrain boundary alongside the robot; the simulation comparison as reported counts only the robot reaching the boundary, so it does not directly verify ball transport.","Because DSF-PO only changes the loss objective, it could be transplanted to other hierarchical skill-switching tasks with a discrete selector and continuous command heads, even outside legged locomotion; the paper does not test this transfer.","The reported strategy for stair descents — kicking the ball down before following — suggests the high-level policy learns a tacit sequence of manipulation-then-locomotion; a natural extension would randomize the ball's starting position and test whether this sequencing persists.","On downhill terrain the ball's velocity is heavily influenced by gravity rather than by robot contact, so a follow-up ablation separating robot-induced ball motion from terrain-induced motion would clarify how much of the observed performance is active manipulation versus passive rolling."],"forward_implications":["If the hierarchical decomposition is correct, long-horizon loco-manipulation problems can be solved by training reusable low-level skills once and learning only a switch policy on top.","Because DSF-PO is a loss-only change, it should accelerate learning in any RL setting where a discrete selector gating continuous commands makes much of the action space irrelevant at any given step.","The zero-shot real-world transfer reported here implies that a simulator-trained skill library plus an onboard ball detector may be enough to deploy ball manipulation on terrain types not seen during training.","The reported success on stair descent and gravel suggests that legged object transport is not limited to flat indoor floors, which would matter for deliveries, search-and-rescue, and robot soccer on outdoor fields."],"supporting_citations":[{"why":"Supplies the prior learning-based dribbling method that this work extends and compares against in simulation and real-world results.","marker":"[5]"},{"why":"Supplies the prior dribbling method with feedback control reward that this work builds on and compares against, including real-world success rates.","marker":"[6]"},{"why":"Inspires one of the low-level locomotion skills used for fast movement on relatively flat terrain.","marker":"[14]"},{"why":"Inspires the rugged-terrain locomotion skill used by the high-level policy for rough terrain traversal.","marker":"[34]"},{"why":"Provides the PPO update algorithm on which both low-level skills and the high-level actor are trained, and which DSF-PO modifies.","marker":"[35]"},{"why":"Provides the asymmetric actor-critic architecture used to train the high-level actor with access to full state in the critic.","marker":"[37]"},{"why":"Provides the massively parallel simulation environment used to train all policies in the paper.","marker":"[12]"},{"why":"Provides the onboard ball detector used for ball localization during real-world deployment.","marker":"[38]"}],"fun_headline_variants":["Hierarchical RL enables quadruped dribbling on rugged terrain","Skill-switching policy teaches a robot to dribble on rough ground","Quadruped dribbles a ball across ramps and stairs","Rugged-terrain dribbling with a two-level policy"],"cache_read_input_tokens":13696,"weakest_assumption_plain":"The load-bearing premise is that the success metric in the main simulation comparison — the robot reaching the terrain boundary — is a valid measure of ball manipulation, even though it does not require the ball to be transported or controlled.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical RL enables quadruped dribbling on rugged terrain","Skill-switching policy teaches a robot to dribble on rough ground","Quadruped dribbles a ball across ramps and stairs","Rugged-terrain dribbling with a two-level policy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3211,"prompt_tokens":917,"completion_tokens":2294,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2223}},"tokens_in":533,"tokens_out":2294,"duration_ms":15112,"temperature":1.0,"reasoning_tokens":2223,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:36:08.505114+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the terrain traversability evaluation with success redefined as the ball crossing the same boundary point as the robot, and count episodes in which the ball is abandoned; if the reported advantage over the baselines disappears under this ball-focused criterion, the central claim of superior ball manipulation is not supported.","supporting_citations":[{"cited_title":"Dribblebot: Dynamic legged manipulation in the wild,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior learning-based dribbling method that this work extends and compares against in simulation and real-world results."},{"cited_title":"Dexdribbler: Learning dexterous soccer manipulation via dynamic supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the prior dribbling method with feedback control reward that this work builds on and compares against, including real-world success rates."},{"cited_title":"Walk these ways: Tuning robot control for generalization with multiplicity of behavior,","cited_arxiv_id":null,"evidence_quote":"Inspires one of the low-level locomotion skills used for fast movement on relatively flat terrain."},{"cited_title":"SLR: Learning quadruped locomotion without privileged information,","cited_arxiv_id":null,"evidence_quote":"Inspires the rugged-terrain locomotion skill used by the high-level policy for rough terrain traversal."},{"cited_title":"Asymmetric actor critic for image-based robot learning,","cited_arxiv_id":null,"evidence_quote":"Provides the asymmetric actor-critic architecture used to train the high-level actor with access to full state in the critic."},{"cited_title":"Learning to walk in minutes using massively parallel deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Provides the massively parallel simulation environment used to train all policies in the paper."}],"review_version":1}