PaperlayerView exact revision on arXiv

Source-linked AI summary

HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training

Xihan Ma, Bo Zhang, Boxiong Li, Yuhan Bao, Ziyue Liu, Yucheng Wang

arXiv:2610.32909v1cs.RO

TL;DR

Humanoid boxing requires task decisions as well as motor execution, but HoloBoxing studies adaptation when a pretrained whole-body controller is already available. It reuses that controller and a compact target decoder, initializes a planner with measured-state imitation, and refines it through reinforcement learning. The seven-minute version produces scored interaction and ranks second in the evaluated pool, while the longer version ranks first.

  • Problem

    Humanoid boxing combines motor execution with opponent-responsive decisions, and the paper asks how much task adaptation is needed when a general-purpose whole-body controller is already available.

    Source: §2 · p. 2, §1 · p. 2

  • Method

    HoloBoxing uses a compact target decoder and frozen whole-body controller, initializes a feedback planner with measured-state imitation, then refines task behavior through reinforcement learning.

    Source: §1 · p. 2, §3.1 · p. 3, §6 · p. 8

  • Results

    The seven-minute training path produces scored boxing interaction and ranks second in the evaluated pool, while the 21.4-minute version ranks first.

    Source: §6 · p. 9, §7 · p. 9

  • Takeaways & Limitations

    The experiments support a recipe that reuses motor execution and target structure while evaluating the behavior obtained within a short task-adaptation budget.

    Source: §6 · pp. 8–9

    The study uses a small motion corpus and one selected task-training seed, and its reported training time excludes controller pretraining, preparation, and search.

    Source: §5.2 · p. 5, §6 · p. 9

Abstract

from arXiv · show

HoloBoxing presents a simple recipe for learning humanoid boxing by building on a pretrained whole-body controller. The recipe separates reusable motor skills from task decisions: a compact planner requests future body-point targets, while the frozen controller handles joint-level execution. A learned decoder maps 32-dimensional actions to structured target windows, and measured-state imitation initializes the planner before reinforcement learning. Task learning then refines interaction behavior while reusing the decoder and controller, avoiding the need to relearn whole-body control. This combination keeps task optimization focused on a compact upper policy and supplies an initial behavior before reward-driven exploration. Starting from that available controller, the recipe accounts for about seven minutes of additional training, including reused representation fitting and a single-GPU task run. This excludes controller pretraining, data preparation, evaluation, and search. Simulated boxing experiments and component controls evaluate the resulting behavior and the roles of imitation initialization and the target representation.

1 Introduction

HoloBoxing separates task decisions from whole-body motor execution by building on a pretrained controller. Its recipe initializes a compact planner with measured-state imitation before reward-driven refinement, then evaluates the resulting behavior and training cost.

1 Introduction

  • Boxing requires more than reproducing a punch: the robot must respond to an opponent and remain upright after exchanges.The paper asks how far a short training recipe can go when whole-body motor capability is already available.Source: §1 · p. 2
  • A frozen pretrained controller executes movement requests while a compact feedback planner selects actions from robot state and opponent positions.A learned decoder maps 32-dimensional actions to short windows of future body-point poses.Source: §1 · p. 2
  • Measured-state imitation initializes the planner before reinforcement learning adjusts its behavior through warmup and opponent interaction.The decoder and lower controller remain reused throughout these stages.Source: §1 · p. 2
  • Controller reuse, a 32-dimensional target representation instead of 225 coordinates, and imitation initialization reduce what task learning must acquire.The initial behavior is available before reward-driven exploration.Source: §1 · p. 2
  • The primary recipe accounts for 7.0 minutes of additional training, while a 21.4-minute version provides a higher-budget comparison.Simulation experiments assess interactive boxing and controls remove imitation initialization or the target decoder.Source: §1 · p. 2

2 Background and motivation

The paper frames boxing adaptation as learning opponent-responsive task decisions above reusable whole-body motor capability. It combines compact target requests with measured-state imitation and reinforcement learning, keeping task adaptation within a feedback planner and reusable execution interface.

2 Background and motivation

  • Boxing combines motor execution with decisions about approaching, striking, and responding to an opponent.The paper focuses on additional learning when a general-purpose controller already supplies whole-body movement.Source: §2 · p. 2
  • HoloBoxing uses the released HoloMotion v2 controller for joint-level execution, so the boxing planner learns task decisions rather than relearning motor control.The reported task-training cost excludes the inherited controller’s pretraining.Source: §2 · p. 2
  • A decoder maps normalized 32-dimensional actions to 225 future body-point target coordinates, giving the planner a compact interface while retaining the lower controller.Target reconstruction quality alone does not establish physical executability, so the paper also evaluates tracking behavior.Source: §2 · p. 2
  • The planner is initialized from measured robot observations collected during reference tracking, then refined through standing warmup and interaction rewards.The decoder and whole-body controller remain frozen during task learning; measured states connect initialization to the feedback available during execution.Source: §2 · p. 3
  • The combined recipe confines task adaptation to a compact feedback planner with initial behavior and a reusable execution interface.The experiments measure training cost and boxing behavior, while component controls examine ingredients without assuming separate causal speedups.Source: §2 · p. 3

3 HoloBoxing

HoloBoxing fits a compact target decoder, initializes a feedback planner from measured-state imitation, and refines task decisions through reinforcement learning while preserving the frozen controller and decoder.

3 HoloBoxing

  • A 1095-dimensional masked planner input retains current-self and opponent-position features, and its 32-dimensional action is decoded into 225 coordinates for five body points across five future times.The frozen controller combines decoded targets with robot feedback to execute movement requests.Source: §3 · p. 3

3.1 Compress movement requests before task learning

  • The first stage fits an encoder and decoder on normalized reference target windows, then freezes the selected decoder and its statistics for task learning.Only the decoder enters the executed boxing policy; the encoder remains an offline fitting and diagnostic component.Source: §3.1 · p. 3

3.2 Initialize feedback before reward-driven exploration

  • The planner learns from measured robot observations and future-target labels collected during reference tracking, with supervision expressed in the measured robot’s feedback frame.Future reference states provide labels rather than actor inputs, and the selected planner weights initialize task learning before reward-driven exploration.Source: §3.2 · p. 4

3.3 Refine task decisions while retaining motor execution

  • PPO uses standing warmup to initialize interaction updates, while the target decoder and HoloMotion controller stay frozen as task learning refines decisions.The primary learner uses a fixed external standing opponent; its reward combines task shaping and adversarial motion style with weights 0.8 and 0.2.Source: §3.3 · p. 4

4 Evaluating short task adaptation

The evaluation measures whether the reused hierarchy produces interactive boxing after the short training procedure. It assesses adaptation time, competitive behavior, and component effects under a shared simulation setup, while noting that observation definitions differ across policies.

4 Evaluating short task adaptation

  • The evaluation measures training time, competitive behavior, and effects of removing task-policy components as distinct aspects of the recipe.Time quantifies adaptation cost, while behavior tests what that budget achieves.Source: §4 · p. 4
  • All policies share the robot, actuator settings, arena, simulator, and referee, but retain different observation definitions, so the league is not observation-controlled.A valid hand contact scores a point, and a sustained fall can end a bout with a knockout.Source: §4 · p. 4
  • A fixed ten-round league gives policies the same opponent coverage with randomized sides and initial-state seeds, and Elo summarizes outcomes within that frozen participant pool.Recorded states, actions, and referee events allow replay without rerunning simulation.Source: §4 · p. 4

5 Experiments: training cost and resulting behavior

The experiments assess a task learner built above a frozen whole-body controller, accounting for 7.0 minutes of additional training and evaluating boxing behavior and design choices. The 7-minute version ranks second in a fixed six-policy league, while component and representation studies examine the tested recipe and its scope.

5.1 A small task learner above a fixed controller

  • The experiments use the Unitree G1 with the frozen HoloMotion v2 controller and a planner that predicts future body-point targets.The decoder predicts five future poses for the pelvis, wrists, and ankles spanning 20–100 ms.Source: §5.1 · pp. 4–5

5.2 Where the seven minutes are spent

  • 422.2 seconds (7.0 minutes) comprise a 375.0-second new process and 47.2 seconds of previously recorded target-representation fitting.This is accounted training cost, not a fresh uninterrupted run of every stage; controller pretraining, evaluation, and hyperparameter search are excluded.Source: Table 1 · p. 5, §5.2 · p. 5

5.3 Additional training above reusable motor capability

  • The comparison supports minute-scale boxing adaptation from an available general-purpose controller, but not a controlled speedup factor or fastest-method claim.Reported methods differ in inherited skills, hardware, observations, and task objectives.Source: §5.3 · p. 5, Table 2 · p. 6

5.4 Boxing behavior after short adaptation

  • The 7-minute version ranks second in the fixed six-policy league, while the long version ranks first.The primary version wins seven of ten bouts against GGSD and four of ten against RoboStriker; four of its wins contain no scored hand contact.Source: §5.4 · p. 6, Table 3 · p. 7

5.5 Roles of initialization and the target representation

  • Under the tested 128-update final-stage budget, the 7-minute version ranks above both component removals and produces more scored contacts.The direct-target variant changes both action dimension and exploration geometry, so this comparison concerns the complete action-interface choice.Source: Table 3 · p. 7, §5.5 · pp. 7–8, Table 5 · p. 7

5.6 Choosing a compact but executable representation

  • Current-state input with opponent information yields 16.9% lower mean supervised validation loss than ten-frame history with opponent input; selected and uncompressed targets each complete 80 of 81 validation windows.The observation result supports imitation initialization, not faster reinforcement learning; a later candidate lowers reconstruction loss but adds a tracking failure and loses required frame coverage.Source: Table 6 · p. 8, §5.6 · p. 8

6 Discussion

HoloBoxing separates motor execution, target structure, and task decisions so task learning can reuse the decoder and controller. The seven-minute path produces scored boxing interaction, while the study limits what can be concluded about speed and generality.

6 Discussion

  • Reusing the decoder and controller across imitation and reward-driven refinement keeps the task learner focused on the upper policy and its objectives.Task actions replace 225 future-target coordinates with 32 latent variables; recorded training time is mostly spent on standing and interaction.Source: §6 · pp. 8–9
  • The experiments do not isolate a causal speedup for each design choice, and the measured seven-minute cost covers additional training above reusable components.The accounting excludes controller pretraining, preparation, and search.Source: §6 · p. 9
  • The seven-minute path ranks second in the evaluated pool, while the 21.4-minute version ranks first and the seven-minute version outperforms both component removals in league outcomes.The seven-minute version uses 128 final-stage updates.Source: §6 · p. 9
  • Evidence is limited by a small motion corpus, one task-training seed, and simulation-only evaluation; league outcomes do not establish method superiority.Compared systems retain different state encodings, histories, and auxiliary inputs.Source: §6 · p. 9

7 Conclusion

HoloBoxing combines a pretrained controller, compact target representation, imitation, and task reinforcement learning to separate task decisions from motor execution. Its primary recipe takes about seven minutes of additional training and produces scored interaction, while a longer version improves league outcomes at greater cost.

7 Conclusion

  • The recipe reuses the controller and decoder while imitation initializes a compact planner before task reinforcement learning refines interaction.The primary recipe accounts for about seven minutes of additional training above an available controller.Source: §7 · p. 9
  • The seven-minute recipe produces scored boxing interaction, while the 21.4-minute version improves league outcomes at greater training cost.Source: §7 · p. 9
  • The experiments support the combined recipe under measured conditions, but do not isolate each component’s contribution to training speed.Source: §7 · p. 9

A Reproduction settings

The reproduction settings describe motion and planner data splits, training configurations, target-coordinate alignment, and reconstruction-study scope. The reported timing includes prior fitting and the new single-GPU process, while excluding preparation and evaluation.

A Reproduction settings

  • The selected setup uses Unitree G1 with the frozen HoloMotion v2 controller and keeps related motion windows within the same split.The prior uses 21,148 training, 11,074 validation, and 3,246 test windows; planner initialization uses 2,568 training and 6,448 validation examples.Source: A Reproduction settings · p. 12
  • The seven-minute accounting includes previously recorded prior fitting and the complete new single-RTX-4090 process, but excludes controller pretraining, preparation, evaluation, and search.The primary version and component removals use seed 43 and the same frozen external standing opponent.Source: A Reproduction settings · p. 12, Table A.2 · p. 12

B Objectives and target coordinates

  • For five points and five future samples, target windows contain 225 coordinates; measured-state imitation uses actual robot observations with future reference states as labels.A shared horizontal alignment translation shifts future points and participants, while orientations remain unchanged.Source: B Objectives and target coordinates · p. 13
  • The target loss uses positive standard-deviation floors, and gradients update the planner while preserving decoder weights; PPO uses correlated latent-action exploration.Exploration scales are 0.1 for standing warmup and 0.2 for task refinement, with temporal correlation 0.9.Source: B Objectives and target coordinates · p. 13

C.1 Observation and reconstruction studies

  • The remaining studies examine decoder capacity and target coverage; the capacity study uses three seeds, and Table C.1 reports dimensionless normalized error with sample standard deviation.Source: §C.1 · p. 13, Table C.1 · p. 13

C.2 Physical tracking and held-out coverage

The selected decoder and uncompressed targets both physically track 80 of 81 validation windows, failing the same source window. Held-out evaluation also exposes a limitation for seated-retreat motion because its test windows begin below the upright-height gate.

C.2 Physical tracking and held-out coverage

  • 80 of 81 validation windows are completed by both the selected reconstruction and uncompressed targets, with the same source window failing.Held-out reconstruction is evaluated after selection, without reselecting the decoder.Source: §C.2 · p. 14
  • A later decoder candidate lowers validation loss from 0.045326 to 0.044545 but adds one tracking failure and misses 16 fixed comparison frames.The selected decoder therefore remains in use.Source: §C.2 · p. 14
  • Table C.2 reports offline reconstruction errors for decoded reference targets before physical execution.Those errors are distinct from the physical tracking evaluation.Source: Table C.2 · p. 14
  • One seated-retreat InterHuman test clip contributes 97.7% of that source’s test reconstruction loss, and its two execution windows start below the upright-height gate.This gate cannot establish whether reconstruction worsens execution of the intended seated motion, limiting claims about representation coverage beyond upright behavior.Source: §C.2 · p. 14

C.3 Reference-to-measured target transfer

The reference-to-measured comparison uses the same source frames and participants with representation weights and statistics fixed. Additional controls assess latent fitting and calibrated physical noise for direct-target exploration.

C.3 Reference-to-measured target transfer

  • The paired source and measured labels use identical source frames and participants, with representation weights and statistics fixed.The comparison includes the coordinate change and alignment translation defined for the transfer.Source: §C.3 · p. 14, Table C.3 · p. 14
  • Latent-only fitting uses 256 evenly spaced measured-validation examples and optimizes only latent codes; its residual is an attained local-fit error, not an expressiveness bound or initialized-planner result.The fitting uses 128 examples per measured validation group and 100 Adam steps at learning rate 0.03.Source: §C.3 · p. 14

C.4 Additional controls

  • 18.26 and 18.34 cm root-mean-square position perturbations are induced by latent and direct-target policies, respectively, at the final-stage setting.Their rotation perturbations are 0.713 and 0.701 radians, respectively; comparable magnitudes do not establish identical covariance or optimization geometry.Source: §C.4 · p. 14

D League rules and opponent interfaces

The league uses a shared simulated setup and fixed rules to compare current participants and two external checkpoints. The section also reports a bounded inference-timing measurement and recorded bout examples, which illustrate behavior rather than estimate win rate.

D League rules and opponent interfaces

  • All six current participants share the packaged robot, actuator parameters, arena, simulator, and contact referee; only the current ten-round, 150-bout league contributes to the reported ranking.Each round covers all fifteen unordered pairs once, with randomized sides and recorded initial-state perturbations and sampling seeds.Source: D League rules and opponent interfaces · p. 15
  • Legal hand contacts to the opponent’s head or torso score at a 5 N·s threshold within 20 ms, while a continuous three-second fall confirms a knockout.A confirmed knockout awards the opponent 1000 points; otherwise bouts use the 60-second timeout score, with equal scores drawing.Source: D League rules and opponent interfaces · p. 15
  • The RoboStriker G1 Stage-3 and GGSD run2 external checkpoints retain source-trained weights and receive no destination fine-tuning.Source: D League rules and opponent interfaces · p. 15

D.1 Inference timing

  • 3.182 ms average and 3.373 ms empirical 99th-percentile latency are measured in a fixed-pose RTX 4090 microbenchmark of the long version.The benchmark excludes physics and output sanitation; fixed-pose samples do not establish full-match or hard real-time latency.Source: §D.1 · p. 15

E Recorded boxing examples

  • Recorded examples show a primary 7-minute-version win against GGSD and a loss against RoboStriker, illustrating approach and scored exchanges in both directions.The examples are reconstructed from saved poses without policy inference or physics stepping and are not an estimate of win rate.Source: E Recorded boxing examples · p. 16, Figure E.1 · p. 16
Loading 2610.32909v1…