acceptodds
Withdrawn from ICLR 2027

BadmintonLab: A Unified Physics-Based Platform for Humanoid Badminton Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent advances in embodied intelligence are pushing humanoid robots beyond locomotion toward dynamic tasks that require active interaction with the physical world. Racket sports provide a particularly demanding setting, where control must be concentrated within a brief window of precise contact, while the effectiveness of each stroke becomes evident only from the ensuing trajectory. Badminton makes these challenges especially pronounced due to the fast and highly dynamic flight of the shuttlecock and the stringent spatiotemporal precision required for racket control. We introduce BadmintonLab, a unified physics-based platform that provides three progressively richer learning settings for task-oriented skill learning, reference-guided style learning, and competitive rally learning. To enable effective learning across these settings, the platform employs compact event-driven task objectives that provide efficient learning signals for racket–shuttle interaction and shot outcomes. Experiments demonstrate distinct badminton skills, reference-guided motion styles, and competitive rally learning, and further show reuse of the same task and learning interfaces across different humanoid embodiments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Humanoid Robotics, Embodied Intelligence, Reinforcement Learning, Physics-Based Simulation, Badminton

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_uB8UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents BadmintonLab, an Isaac Lab-based simulation platform for humanoid badminton. It has three settings built on a shared substrate (court, drag-based shuttle model, rigid-body racket contact, rule checking, and a shooting solver for incoming shuttles):

Level 1, task-oriented skill learning. A single humanoid learns clear, drop, and lift without reference motions. Each skill is a task specification (D_in, H, G): the incoming-shuttle distribution, the interception-height interval, and the target landing region. Level 2, reference-guided style learning. An AMP motion prior is added on top of the Level-1 objectives, using retargeted Unity-asset motions. Level 3, competitive rally learning. Two humanoids sharing one policy play self-play with a gated win/loss reward. The main technical idea is a set of "compact event-driven task objectives": a racket–shuttle approach term evaluated as a minimum distance inside a receive interval, followed by contact, outgoing-shot, net-clearance, and landing terms. The authors argue that this removes the need for a commanded time-to-hit or time-indexed swing phase, in contrast to prior humanoid or robotic badminton systems.

Validation is entirely in simulation:

T1, Level 1 and 2 (Table 1). Hit rates are about 100% and valid-return rates are 89–100%. Target-landing rates are 56–82%. Style (Figure 3). A PCA plot of transition embeddings shows Level-2 transitions shifting toward the reference distribution. Self-play (Figure 4). A checkpoint pairwise win-rate matrix and landing statistics from 30k self-play iterations. Cross-embodiment (Table 1). T2 results are reported for Level-2 clear and lift.

Critical strengths
  1. Timely problem with a reusable interface. Humanoid racket-sport learning is a relevant frontier, and a reproducible simulation setting for it would be valuable. The (D_in, H, G) task parameterization and the event-based reward interface are a clean design. The paper shows that one reward template yields three different strokes by changing only the task specification. The shooting solver and the evaluation protocol (409,600 attempts per policy) are carefully documented. This is the main reason I see potential value if the artifact is released and validated.

  2. A physics-coupled two-humanoid rally setting. Level 3 couples two whole-body humanoid policies through racket–shuttle contact, shuttle flight, and court rules. This is rarely available in prior robotic badminton work, which focuses on single-agent stroke acquisition. The result-reward gating by both players' valid-return rates is a sensible design for avoiding premature terminal signals. Even though the current evidence is thin (see below), the setting itself is a worthwhile contribution.

Critical weaknesses
  1. The central methodological claim has no controlled evidence. The paper's main distinction from Ma et al., Liu et al., and Chen et al. is that event-driven objectives need no commanded contact time and avoid the pitfalls of long-horizon proximity rewards. This is never tested. There is no baseline using a time-gated objective in the same simulator, and no ablation of the interval-limited approach term. For example, the paper says a continuous proximity reward "may encourage the racket to remain close to the shuttle rather than strike actively" (Sec. 3.2), but never tests that hypothesis. There are also no ablations of the outgoing-shot, net-clearance, or landing terms. The evidence for "active striking" is a single T1 rollout (Figure 2), which cannot support a general claim.

  2. The claim is also overstated conceptually. The receive interval I_H(τ) is defined by the shuttle descending through the task-given height interval H, and the shuttle state is observed. The window therefore carries the same timing information implicitly, and the policy can infer time-to-hit from the shuttle state. The approach reward is also evaluated once when the interval ends, so it is still temporally gated in effect. The paper does not discuss this. Comparisons with the prior systems are made only in prose, based on the authors' reading of their reward structure, with no numerical comparison. As a result, the paper's novelty rests on an untested design choice combined with established components (PPO, AMP, Isaac Lab, OmniRetarget/GMR), and the "platform" contribution is not demonstrated as enabling anything beyond what a single configuration achieves.

The experimental evidence is statistically thin and in places over-interpreted.

No seed variance. Each setting appears to be a single trained policy. The 409,600 evaluation attempts reduce evaluation noise but say nothing about training variance. Differences such as the Level-1 to Level-2 changes (lift R_target 64.65% to 56.24%, described as "largely preserved") cannot be interpreted without multiple seeds. Level 3 conclusions are weaker than described. The headline is that every checkpoint beats the 0k policy (53.9–94.1%). This is the easiest comparison, since 0k is only a stroke-capable initialization. Among trained checkpoints the matrix is non-transitive and non-monotonic. For example, 10k beats 5k at 80.3%, but 10k wins only 30.3% against 30k. 5k wins only 32.6% against 20k. Beyond 10k, most pairs sit in the 40–60% range. No confidence intervals are given. The "distance to losing opponent" trace is also not monotone (0.938, 1.279, 1.707, 1.163, 1.951, 1.623 m for 5k to 30k). The text reports an increase from 1.03 to 1.62 m by comparing endpoints and does not mention the 25k value of 1.951 m or the intermediate dips. Deeper returns are not surprising. The text presents deeper returns as emergent because placement is not rewarded. They follow directly from a +10 win/loss reward in a setting where landing in the opponent's half decides the rally. Self-play may not be competitive. Both agents use one shared policy with mirrored observations. No external opponent (heuristic, independently trained, or best-response) is evaluated, and no rally-length, error-type, or exploitability statistics are given. The results may therefore reflect cooperative long exchanges rather than competitive strategy, and the paper does not distinguish these. Style evidence is weak. It consists of one PCA projection that explains about 26% of the variance (17.5% plus 8.7%). There is no quantitative distribution distance in the full feature space. The Level-2 setup also removes root-orientation and default-posture regularizers (Table 4), so the Level-1 to Level-2 difference may come from this change rather than the AMP prior, and no control isolates it. The reference set (locomotion, preparation, receive, serve) does not clearly include stroke swings, so the claim about shaping the "style of strokes" needs qualification. The T2 results are unexplained. T2 Level-2 reaches R_target of 99.2% (clear) and 99.5% (lift), versus 82.48% and 56.24% for T1 on the same tasks. This is a very large gap on the hardest skill. Section 4.4 says only "high target-placement success". Without matching configurations, T2 Level 1, and an explanation, this undermines the "cross-embodiment reuse" claim. It is unclear whether the tasks are comparable or whether T2 is simply an easier embodiment. e_land is weakly informative. The target regions are large and error is zero inside them (e.g., clear G is 0.7 m by 4.88 m). A small mean error over valid returns, with valid-return rates as low as 89%, says little about placement quality. No dispersion of landing points or reference baseline is given.

Questions for authors
  1. Can you implement a time-gated baseline in your simulator (commanded time-to-hit with a swing reward near that time) and compare it with the event-driven objective on hit rate, target rate, sample efficiency, and the distribution of contact time and height? Can you also ablate (a) the interval restriction in d_app, (b) each of the outgoing, net, and landing terms? Without these, I cannot credit the central design claim.
  2. Can you report results over multiple training seeds with means and standard deviations for Table 1 and Figure 4, with confidence intervals on win rates? In particular, is the lift decline from Level 1 to Level 2 outside seed variance?
  3. Why does T2 do so much better than T1 (99.5% vs 56.24% on Level-2 lift)? Are the task specifications, rewards, and training budgets identical? Please provide T2 Level 1 and drop results.
  4. In Level 3, how does the policy fare against external opponents (an independently trained policy, a scripted opponent, an earlier-seed policy) and under a best-response test? What are the average rally lengths and the fault-type breakdown? Does the shared-policy self-play learn to compete, or to sustain long exchanges?
  5. Does the policy implicitly infer time-to-hit from the observed shuttle state? How does performance change with observation noise or delay?
  6. For the style claim, can you provide a quantitative distribution metric in the full feature space and a control that removes the Level-2 regularizer changes without AMP? Does the reference set contain stroke swings?
  7. Will the code, task configurations, trained checkpoints, and reference-motion data be released, and is redistribution of the Unity assets permitted?
Reviewer_tfarOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

BadmintonLab is a simulation platform for humanoid badminton learning. It attaches rewards to physical events rather than to a commanded swing: the approach term is the closest racket–shuttle distance inside an interception-height window, and outgoing speed, net clearance and landing are scored after contact. Three levels share the interface: per-skill clear/drop/lift policies, AMP-based style learning, and two-humanoid self-play with only the approach term and a win/loss reward. Experiments on T1 and T2 show reliable returns, reference-shifted motion in Level 2, and deeper returns after self-play.

Critical strengths
  1. Rewards are tied to the return outcome, not to the swing: the approach term constrains only the interception-height window, while outgoing speed and direction, net clearance and landing enter the reward directly (Sec. 3.2, Eq. 8–12; Table 3).

  2. Level 3 connects win/loss feedback to whole-body control. Self-play is not new (Haarnoja et al., 2024, 1v1 soccer), but each badminton exchange is a discrete strike with a delayed outcome; the rally state machine and win/loss reward link the physical and rally layers (App. B.4, C.2).

Critical weaknesses
  1. None of the three levels is framed as a testable question. Figure 2's swing comes from the full reward configuration, which includes an outgoing-shot term targeting 10 m/s while the shuttle falls at 5–7 m/s, so the swing cannot be attributed to the window objective; H constrains only the approach term (App. B.1), and no contact-height distribution is reported. There is no comparison with alternative rewards and no ablation. This decides whether the main claim holds.

  2. The gain from Level 3 cannot be attributed to the win/loss feedback. Table 10 shows the reward was added while the outgoing, net-clearance and landing shaping were removed, so the deeper returns may be drift. The missing control is to start from 0k, remove only the win/loss reward, and compare against the same fixed opponents. The win-rate matrix supports only "later checkpoints beat 0k": row means excluding 0k lie in 45.6–54.6%, with no monotone trend and clear intransitivity. The unit of W = 8 and the 0k interface (104 → 120 dimensions, no racket terms, window [1.0, 1.5] m) are also undefined.

Questions for authors
  1. Both sides share one policy (App. C.2), so the opponent is the current policy. Has this self-play been compared with a frozen snapshot pool or with fixed opponents? How are the first and fifth shots of a rally credited differently under a terminal ±1? Please add one ablation: win/loss versus approach-only for the same number of steps.

  2. What problem is AMP meant to solve? Since root orientation and default-posture matching were removed at the same time (Table 4) and the lift target rate falls from 64.65% to 56.24%, is this due to AMP or to the removal? Please give both controls, three seeds each. Why AMP rather than tracking or an explicit style reward?

  3. Eq. 8 takes a minimum per receive interval with a 0.10 m full-reward radius and a decay ρ whose scale is not given: what is that scale, and what does the term shape if it is much wider than 0.10 m? All four event terms carry weight 1.0 although their scales differ; was any sensitivity tested? If contact occurs before the shuttle enters H, how is the approach term settled?

  4. §4.1 trains a separate policy per skill. If (D_in, H, G) fully describes the task, why not one conditioned policy? Since clear and drop differ only in the target region, does the policy learn a shot type or just "send the shuttle to a region"?

Reviewer_41JYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces BadmintonLab, a collection of physics-based RL training setups for humanoid badminton. It explores three settings: (1) task-oriented skill learning using PPO with manually designed rewards, (2) reference-guided style learning using AMP, and (3) competitive rally learning through self-play. The training and evaluation is in Isaac Lab, and are primarily conducted on Booster T1, with additional evaluations on Booster T2.

Critical strengths

The paper introduces rewards that allow policies to learn when to strike without prescribed contact timing, and extends humanoid badminton to competitive self-play. While these ideas are common in other domains, their application to humanoid badminton is relatively new.

Critical weaknesses

The work mainly combines existing PPO, AMP, and self-play techniques without substantial algorithmic contributions. It overlooks important related work, including Learning Physically Simulated Tennis Skills from Broadcast Videos (SIGGRAPH 2023), Strategy and Skill Learning for Physics-based Table Tennis Animation (SIGGRAPH 2024), and SMPLOlympics (2024), which already explore task rewards, reference-guided style learning, and competitive sports.

The three levels appear largely independent, with no clear learning progression between them. And each skill requires a separately trained policy, while Level 2 only demonstrates two skills without diverse or controllable motion styles.

The paper lacks comparisons with existing badminton methods and ablations validating the proposed rewards.

Level 3 mainly compares checkpoints from the same self-play training, with limited evidence of meaningful strategic learning.

Questions for authors

Are the three levels independent or sequential? Is Level 2 initialized from Level 1, and which policy initializes Level 3?

Why does Level 2 only demonstrate two skills, and where is the evidence of learning diverse motion styles? Why are different skills trained separately rather than using a unified multi-skill policy?

What meaningful competitive strategies emerge from Level 3 beyond deeper shot placement?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.