{"id":"d313162b-63ee-4141-a237-a59500666fa4","arxiv_id":"2507.11498","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A simulated Unitree G1 humanoid learns to drum dozens of popular songs from MIDI with high F1 scores using a Rhythmic Contact Chain and temporal decomposition.","lead":"This paper trains a simulated humanoid robot to play drums using reinforcement learning, converting MIDI drum tracks into timed contact goals. It reports high accuracy across 32 popular songs and shows that dense contact rewards and goal conditioning are essential.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"F1 scores hinge on an undefined hit-to-target matching rule; without a stated timing tolerance or onset-asynchrony distribution, the headline claim of rhythmic precision cannot be verified.","rationale":"The reader's weakest-assumption analysis correctly identifies the undefined timing tolerance in the F1 metric as the key ambiguity. In good faith, the paper is a solid empirical study with a clear pipeline, useful ablations, and a plausible central mechanism: dense contact rewards plus future one-hot contact goals enable a learned policy to follow a rhythmic contact chain. The reward ablation and the near-zero F1 without contact goals provide credible internal support. However, the headline numbers are only meaningful if the hit-to-target matching rule is fixed. The paper states that F1 compares 'actual hits at each timestep,' which suggests a strict 20 ms tolerance, but it never specifies the matching procedure, duplicate handling, or the force threshold that defines a registered hit. This is not an external consensus disagreement; it is an internal evaluation-validity gap. The concrete test would settle whether the scores survive an explicit, tolerant matching rule. I therefore do not move the verdict: the appropriate disposition remains conditional on the authors specifying and releasing the evaluation protocol, and on verifying that the reported F1 values are robust to a reasonable timing tolerance.","tokens_in":13167,"tokens_out":4215,"duration_ms":61549,"concrete_test":"Obtain or reconstruct the evaluation script, then recompute F1 under an explicit event-matching rule: associate each target onset to the nearest actual strike on the same drum using one-to-one assignment, with tolerance τ ∈ {0, 20, 50, 100} ms, and report F1 for each τ plus the distribution of onset asynchronies. Also cluster contact events so that one physical strike spanning multiple 50 Hz timesteps counts once. If F1 changes by more than 0.05 between τ=0 and τ=50 ms, or if duplicate-contact suppression materially changes scores, the reported F1 values are artifacts of an unstated matching rule and the central precision claim is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a PPO policy achieves high F1 scores (mostly above 0.9) across 32 tracks, with the implication of split-second timing on a humanoid drummer. The load-bearing condition is that the F1 metric and the dense contact reward in §III-D and §IV-B unambiguously measure temporally correct strikes. The paper never defines how an actual drum hit is matched to a target contact event. It only says scores are computed by comparing the ground-truth RCC to 'the robot's actual hits at each timestep.' With a 50 Hz control timestep, a strict per-timestep match makes the implicit tolerance exactly 20 ms: a hit one timestep early or late is counted as a wrong or missed strike. This is an unusually strict and consequential definition, and the paper gives no evidence that the metric is implemented this way, nor any alternative matching rule. The problem is compounded by contact detection: Isaac Gym registers an impact when a collision body is struck with 'sufficient force,' but the threshold is not reported, and a single physical strike may register over multiple timesteps. Without the evaluation script, the reported F1 scores are not independently interpretable: the same policy could score 0.9 or 0.6 depending on the matching rule and contact-duplicate handling. Because the paper's novelty is precisely rhythmic precision, this undefined metric is the most load-bearing weakness in the empirical argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a simulation framework, Robot Drummer, in which a Unitree G1 humanoid learns to play MIDI-derived drum tracks by treating drumming as a sequence of timed contact events, the Rhythmic Contact Chain (RCC). The RCC is segmented into fixed-length intervals, and a single PPO policy is trained across all segments in parallel with dense contact rewards (correct, wrong, missed, proximity) plus a lookahead of future contact goals. The authors report specialist policies achieving mean F1 scores mostly above 0.9 across 32 rock and metal songs, present ablations supporting each reward term and the contact-goal observation, show that a single generalist policy over multiple songs degrades performance, and include a small listener study. The overarching claim is that RL can produce temporally precise, expressive humanoid drumming in simulation.","tokens_in":13491,"tokens_out":4552,"duration_ms":54918,"significance":"If the reported F1 scores are trustworthy, the paper advances process-driven whole-body control: it demonstrates that a single policy can sustain long-horizon rhythmic contact sequences, with emergent behaviors such as cross-arm strikes and adaptive stick assignment. The strengths include a clear contact-chain formulation, a sensible temporal decomposition, hypothesis-driven song-complexity analysis using nPVI and polyphony, and ablations reported as mean ± std over five seeds. The main weakness is that the central metric — F1 against the RCC — is not defined precisely enough to support the claimed rhythmic precision; the paper needs to specify the hit-to-target matching rule, the contact-detection threshold, and a few key hyperparameters before the empirical contribution can be fully assessed.","major_comments":[{"comment":"The F1 evaluation metric is not sufficiently defined. The text says scores are computed by comparing the ground-truth RCC to \"the robot's actual hits at each timestep,\" but no matching rule is stated. With the 50 fps discretization mentioned in the Figure 7 caption, a strict per-timestep match would imply a 20 ms timing tolerance; any hit one timestep early or late would count as wrong or missed. If instead a tolerance window is used, its size and alignment procedure must be reported. Without this information, the reported F1 scores are not interpretable, and the headline claim of near-perfect rhythmic precision cannot be verified. Please specify the matching rule, report the onset-asynchrony distribution for representative tracks, and ideally release the evaluation script.","section":"§IV-B and Figure 7"},{"comment":"The contact-event detection underlying both the reward and the metric is underspecified. The paper states that each drum \"registers an impact when struck with a sufficient force,\" but the force threshold is not reported, and the text does not explain how the set E_t of drums actually hit at time t is formed. A single physical strike could plausibly register over multiple simulation timesteps, which would inflate correct-strike counts and distort precision and recall. Please report the contact threshold, any debouncing or duplicate-removal procedure, and the raw confusion-matrix counts for at least one song.","section":"§III (opening paragraph) and Table I"},{"comment":"Several quantities that determine the training setup are missing. Equation (3) defines the segment length P as N/M, but the actual value of P (or M) is never given; the lookahead horizon L in Section III-D.1 is not reported; and in Equation (4), the action scaling β and the PD gains Kp and Kd are not specified. These parameters directly affect both sample efficiency and the timing precision the policy can achieve, so their absence makes the experiments unreproducible and weakens the \"efficient learning\" claim. Please add a full hyperparameter table.","section":"§III-C and §III-D"}],"minor_comments":[{"comment":"The decision to keep only the most frequent articulation of each drum is justified by saying evaluation \"showed no significant drop in musical quality,\" but no quantitative evidence is provided; consider reporting the affected tracks or a small ablation.","section":"§III-A"},{"comment":"The checkmark notation in the observation-ablation rows is difficult to parse: the row label \"Proprioception & Spatial Phase Variable Contact Goals\" with three checkmarks does not make clear which component is present. Use explicit column headers or a clearer legend.","section":"Table III"},{"comment":"The listener study is described as \"small-scale,\" but the report would benefit from stating the number of ratings per video, whether the 15 participants all rated overlapping sets of videos, and how the Likert scores were aggregated across participants.","section":"§V-D"},{"comment":"The Discussion states that the robot \"kept perfect timing,\" but the highest mean F1 is 0.985 and the paper does not report any direct timing-error statistic. Consider replacing \"perfect timing\" with a qualification such as \"near-perfect per-timestep hit alignment\" or reporting the mean absolute onset error in milliseconds.","section":"§VI"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the undefined evaluation metric in §IV-B: the F1 scores are the main quantitative contribution, and until the matching rule, tolerance, and contact-detection handling are specified, the empirical claims are not fully verifiable. The formulation, ablations, and complexity analysis are otherwise reasonable, so I see this as a fixable major revision rather than a rejection. If the authors provide an evaluation script or detailed metric specification, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on the Robot Drummer paper. The headline: it's the first RL-trained full humanoid drummer, and the core formulation—drumming as a sequence of timed contact events (RCC)—is clean and works. The paper is a solid empirical study, not a toy demonstration.\n\nWhat I like: the ablations are genuinely informative. Removing wrong/missed strike penalties drops F1 substantially, and removing the one-hot contact goals collapses performance. The temporal decomposition is shown to cut training time from ~8 hours to ~2-3 hours with no loss in F1, which is a real practical contribution. The complexity analysis (nPVI, polyphony, drum count) matches intuition and gives the field a useful way to characterize drumming difficulty. The listener study is a nice addition; the low naturalness score (2.46) is honestly reported.\n\nThe main soft spot is the evaluation metric. The paper says F1 compares ground-truth RCC to 'the robot's actual hits at each timestep,' with no tolerance or matching rule. At 50 Hz, that implies a 20 ms window, which is extremely strict. The contact detection threshold in Isaac Gym is also unreported. So the headline F1 values are not portable—another group cannot reproduce them without the exact evaluation script. This is the load-bearing weakness: the paper's claim of split-second timing rests on a metric that is not fully specified. That said, I don't think there's a logical error. The reward uses the same per-timestep matching, so the policy is trained toward the same objective. The numbers are internally consistent. It's a reporting gap, not a fatal flaw.\n\nOther smaller issues: segment length P, lookahead L, PD gains, and action scaling are unlisted, which makes replication harder. The 'emergent human-like strategies' claim is qualitative and the listener study says otherwise on naturalness. But those are minor.\n\nBottom line: this is a worthwhile contribution to robot musicianship and contact-rich RL. It deserves a serious referee, but the metric definition and code release need to be fixed before it can be fully trusted. I'd accept for review.","headline":"First RL-trained humanoid drummer with a clean formulation and honest ablations; the F1 metric needs to be specified before the headline numbers can be trusted.","tokens_in":13989,"tokens_out":2310,"would_cite":true,"duration_ms":27468,"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":"The paper argues that a full drum track can be learned as a Rhythmic Contact Chain and that a single policy, trained on fixed-length segments with dense contact rewards, plays over thirty songs with F1 mostly above 0.9.","keywords":["humanoid control","robotic musicianship","reinforcement learning","rhythmic contact chain","drumming","contact-rich manipulation","temporal decomposition","F1 score"],"falsifier":"Record the actual contact timestamps of a trained policy on a full song, define a hit as correctly matched to a target only when it falls within a stated tolerance window (for example, plus or minus 20 ms), and recompute F1; if the scores drop well below the reported values, the paper's central performance claim depends on the undefined hit-matching rule rather than on genuine rhythmic precision.","tokens_in":12986,"feed_emoji":"🥁","tokens_out":11297,"duration_ms":114916,"temperature":0.7,"pith_summary":"This paper argues that reinforcement learning can take humanoid robots into expressive, process-driven tasks, and uses drumming as the testbed. Its central claim is that a full drum track can be encoded as a Rhythmic Contact Chain---a sequence of timed contact events specifying which drums to strike and when---and that a single policy trained on fixed-length segments can execute that chain over songs several minutes long. Across more than thirty popular rock, metal, and jazz tracks, the trained specialist policies reach mean F1 scores mostly above 0.9, including on demanding drum fills and snare rolls. The learned policies also show strategies the reward never encoded explicitly, such as crossing arms to reach distant drums and reassigning sticks depending on which hit comes next. A sympathetic reader would care because drumming demands split-second timing and sustained multi-limb coordination over a long horizon, the very properties that make process-driven skills hard for current robot learning methods.","feed_headline":"One policy drums 30+ tracks, F1 mostly above 0.9","feed_subtitle":"Dense contact rewards and future-hit goals teach a simulated humanoid to nail fills, rolls, and cross-arm strikes.","key_machinery":"The central object is the Rhythmic Contact Chain (RCC): a sequence of contact steps $C = \\{C_1, \\ldots, C_N\\}$, each step $C_i = (D_i, t_i, S_i)$ specifying the subset of drums to strike at a given time and the sticks allowed to strike them. It converts a MIDI drum track into spatiotemporal targets without prescribing any reference trajectory. Two mechanisms make the chain learnable: the policy observes the next $L$ one-hot contact goals as a discrete phase marker, telling it which drum is coming next, and a dense contact reward scores correct, wrong, and missed strikes plus a proximity bonus at every timestep. Temporal decomposition then divides the chain into fixed-length segments that are trained in parallel by a single policy, turning a several-minute sparse-reward task into sample-efficient reinforcement learning. The RCC carries the argument because it turns aesthetic timing into a dense contact-prediction and contact-execution problem.","core_discovery":"The authors' claim is that a musical performance is, for a robot, a chain of contact-fulfillment events, and that this chain is learnable. They formalize the Rhythmic Contact Chain as $C = \\{C_1, \\ldots, C_N\\}$, with each contact step $C_i = (D_i, t_i, S_i)$ marking which drum(s) must be hit at time $t_i$ and which stick(s) may be used; the chain is converted into one-hot vectors that serve as lookahead goals in the policy's observation. The reward scores every contact event at each control step, rewarding correct strikes, penalizing wrong and missed strikes, and adding a proximity bonus, while temporal decomposition splits each song into fixed-length segments trained in parallel. With this setup, a single specialist policy reproduces hundreds of precisely timed hits per song, scoring F1 mostly above 0.9 across over thirty tracks. The policies also develop emergent human-like drumming strategies, including cross-over strikes and adaptive stick assignment, that were not hand-authored. The paper presents this as evidence that reinforcement learning can bring humanoid robots into creative musical performance.","pith_inferences":["Beyond the paper: if the F1 scores are robust to a defined hit-matching tolerance, the Rhythmic Contact Chain formulation should transfer to other percussion instruments and to any contact-timing task where the goal is a temporal sequence of impacts rather than a final pose.","Beyond the paper: the strong nPVI-to-F1 correlation suggests a testable curriculum---training segments ordered from low to high rhythmic irregularity---that might reduce the specialist-to-generalist performance drop the paper observes.","Beyond the paper: because the paper never defines hit matching, a direct check is to recompute F1 with a one-timestep tolerance and with a 20 ms tolerance; the comparison would show whether the reported high scores reflect true rhythmic precision or a permissive scoring rule."],"forward_implications":["Specialist policies achieve near-perfect rhythmic fidelity on many full songs, with F1 scores above 0.9 on most of the more than thirty tested tracks.","Songs with highly irregular hit timing (high nPVI) or many different drums are the hardest, and polyphony places an upper bound on F1 because the robot can hit at most two drums at once.","A single generalist policy trained across multiple songs performs markedly worse than song-specific specialists, showing negative transfer in multi-song rhythmic control.","Removing any dense-contact reward component or the one-hot contact-goal observations degrades F1 substantially, and contact goals cannot be replaced by a continuous time-phase variable.","Temporal decomposition does not change final F1 but cuts training time from about 8-9 hours to 2-3 hours on the same hardware."],"supporting_citations":[{"why":"supplies the on-policy optimization algorithm used to train all drumming policies.","marker":"[17]"},{"why":"supplies the GPU-parallel physics simulation used for training and evaluation.","marker":"[19]"},{"why":"establishes RL-based piano playing with dexterous hands, the musical-skill baseline this work extends to humanoid drumming.","marker":"[6]"},{"why":"prior RL for a simplified drumming arm, the closest existing drumming result this work generalizes.","marker":"[7]"},{"why":"provides the phase-variable and state-initialization ideas adapted for contact goals and music state initialization.","marker":"[5]"},{"why":"provides the random state-initialization technique used to expose the policy to mid-song phases.","marker":"[18]"},{"why":"introduces sequential-contact decomposition for contact-rich control, motivating the Rhythmic Contact Chain.","marker":"[12]"},{"why":"defines the drum-kit configuration used for training and evaluation.","marker":"[16]"},{"why":"supplies the nPVI rhythm-complexity metric used to characterize songs.","marker":"[20]"}],"fun_headline_variants":["One policy drums 30+ tracks with F1 above 0.9","Robot drummer learns to play 30+ songs via RL","Humanoid robot nails cross-arm drumming on 30 tracks","Drumming robot: single policy, F1 mostly above 0.9","Rhythmic Contact Chain enables long-horizon robot drumming"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported F1 scores count a strike as correct only if it lands on the exact 50 Hz timestep of the intended target, and the paper never defines how a physical hit is matched to a timestep, so the high scores implicitly assert sub-20 ms precision without demonstrating it.","fun_headline_variants_meta":{"raw":{"variants":["One policy drums 30+ tracks with F1 above 0.9","Robot drummer learns to play 30+ songs via RL","Humanoid robot nails cross-arm drumming on 30 tracks","Drumming robot: single policy, F1 mostly above 0.9","Rhythmic Contact Chain enables long-horizon robot drumming"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1519,"prompt_tokens":983,"completion_tokens":536,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":443}},"tokens_in":599,"tokens_out":536,"duration_ms":6220,"temperature":1.0,"reasoning_tokens":443,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:07:24.732638+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the actual contact timestamps of a trained policy on a full song, define a hit as correctly matched to a target only when it falls within a stated tolerance window (for example, plus or minus 20 ms), and recompute F1; if the scores drop well below the reported values, the paper's central performance claim depends on the undefined hit-matching rule rather than on genuine rhythmic precision.","supporting_citations":[{"cited_title":"Isaac gym: High performance gpu based physics simulation for robot learning,","cited_arxiv_id":null,"evidence_quote":"supplies the GPU-parallel physics simulation used for training and evaluation."},{"cited_title":"Robopianist: Dexterous piano playing with deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"establishes RL-based piano playing with dexterous hands, the musical-skill baseline this work extends to humanoid drumming."},{"cited_title":"Em- bodied intelligence for drumming; a reinforcement learning approach to drumming robots,","cited_arxiv_id":null,"evidence_quote":"prior RL for a simplified drumming arm, the closest existing drumming result this work generalizes."},{"cited_title":"Wococo: Learning whole-body humanoid control with sequential contacts,","cited_arxiv_id":null,"evidence_quote":"introduces sequential-contact decomposition for contact-rich control, motivating the Rhythmic Contact Chain."},{"cited_title":"Multisensory Learning Framework for Robot Drumming","cited_arxiv_id":"1907.09775","evidence_quote":"defines the drum-kit configuration used for training and evaluation."},{"cited_title":"The pairwise variability index as a measure of rhythm complexity,","cited_arxiv_id":null,"evidence_quote":"supplies the nPVI rhythm-complexity metric used to characterize songs."}],"review_version":1}