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

cs.ROOct 23, 2026

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

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

Abstract

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.

Figures & tables

Stage	Training or accounting scope	Runtime (s)
Reused target prior	Earlier fitting run; epoch 148 selected	47.2
Imitation initialization	20 epochs; final epoch selected	6.7
Standing warmup	32 updates	84.9
Interaction learning	128 updates	275.3
Other new-process time	Setup, saving, and exit	8.1
Total	Prior fitting plus complete new process	422.2

Table 1: Training-time accounting for HoloBoxing (7-min version). The new process uses one RTX 4090; target-prior fitting is reused from an earlier run.

Method	Existing motor capability	Reported additional training
HoloBoxing	General-purpose HoloMotion v2 controller	7.0 minutes including representation fitting, initialization, and task learning
SLMP	Tracking controller and latent prior	Approximately 8 hours for two-agent combat
RoboStriker	Tracking controller and latent prior	Task-stage wall-clock time unreported
GGSD	Joint skill discovery and gameplay learning	No separately timed post-controller adaptation stage reported

Table 2: Reported additional-training context. Existing motor-controller and prior pretraining are excluded from task-stage times. HoloBoxing additionally includes fitting its target representation and imitation initialization. Unreported denotes no wall-clock duration in the checked sources.

Figure 2: Elo trajectories and final ranking over 150 simulated bouts. Each pair plays ten bouts under the same robot parameters, arena, and referee.

Participant	Elo	W–L–D	Effective contacts
HoloBoxing (long version, 21.4 min)	1669	41–3–6	327:64
HoloBoxing (7-min version)	1541	26–17–7	76:196
RoboStriker	1524	28–22–0	337:29
GGSD run2	1476	20–25–5	69:155
w/o latent decoder (7-min version)	1428	4–22–24	3:165
w/o imitation (7-min version)	1362	4–34–12	0:203

Table 3: Complete-round snapshot of the shared-arena league. Contact counts are candidate-relative totals. The 7-min version and both component removals share a 128-update final-stage budget. The long version uses 256 final updates. HoloBoxing variants are evaluated as distinct league participants.

Opponent family	W–L–D	Contacts	Falls
GGSD	7–3–0	21:16	0:5
RoboStriker	4–6–0	4:85	2:6
Component removals	13–0–7	30:0	0:0
Long version	2–8–0	21:95	0:0

Table 4: Primary 7-min-version results by opponent family in the same complete-round snapshot. Contacts and falls report the primary version first.

Variant	Change from full recipe	New imitation epochs	New run (min)
HoloBoxing (7-min version)	No component removed	20	6.25
w/o imitation (7-min version)	Remove imitation updates	0	5.99
w/o latent decoder (7-min version)	Predict 225 target values directly	20	6.12

Table 5: Single-seed component controls with 128 final updates. New-run minutes include startup, setup, training, saving, and exit. The shared prior is accounted for separately below.

Self observation	Opponent input	Mean validation loss	Sample standard deviation
Ten-frame history	Yes	0.53513	0.00617
Ten-frame history	No	0.54799	0.01025
Three-frame history	Yes	0.46297	0.00333
Three-frame history	No	0.48215	0.00579
Current state	Yes	0.44447	0.00128
Current state	No	0.45782	0.00239

Table 6: Planner observation study. Each condition uses three seeds and the same frozen decoder; values are selected validation losses.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Source	Split	Groups	Files	Duration (s)	Prior targets
RoboStriker	Train	8	16	181.72	18092
InterHuman	Train	6	6	31.16	3056
RoboStriker	Validation	2	4	46.18	4598
InterHuman	Validation	2	2	64.96	6476
RoboStriker	Test	2	4	27.24	2704
InterHuman	Test	2	2	5.62	542

Appendix

Table A.1: Motion coverage by source and split. Duration counts each original source motion once; mirrors and the two participants do not multiply it.

Symbol or component	Meaning	Value
$d_{o}$	Public observation dimension	1097
$K,H$	Body points; future samples per point	5; 5
$d_{p},d_{r}$	Position and orientation feature dimensions	3; 6
$d_{g},d_{z},d_{a}$	Target, latent, and motor-action dimensions	225; 32; 29
$\Delta t$	Control interval	20 ms; targets span 20–100 ms
Prior network	Encoder hidden widths	512, 256; reversed in decoder

Appendix

Table A.2: Implementation settings for the primary 7-min configuration.

Encoder hidden widths	Normalized error	Position error (cm)	Rotation error (degrees)
256, 128	$0.052182\pm 0.001057$	2.653	5.249
512, 256	$0.046047\pm 0.000709$	2.474	4.839
512, 256, 128	$0.055442\pm 0.002174$	2.985	5.092

Appendix

Table C.1: Prior-capacity study. Normalized error is dimensionless; its variation is the sample standard deviation across three seeds.

Split and source	Windows	Normalized error	Position (cm)	Rotation (degrees)
Validation: RoboStriker	4598	0.01985	1.595	3.441
Validation: InterHuman	6476	0.06342	2.902	5.890
Test: RoboStriker	2704	0.00658	0.985	2.269
Test: InterHuman	542	0.78406	6.752	19.055

Appendix

Table C.2: Offline reconstruction of the selected prior. These errors describe decoded reference targets before physical execution.

Target population	Samples	Position error (cm)	Rotation error (degrees)
Paired source-frame references	6448	2.90	5.88
Measured-frame labels	6448	6.30	10.36
Measured subset, encoder initialization	256	6.37	12.27
Same subset, latent-only fitting	256	6.03	8.61

Appendix

Table C.3: Target-interface diagnostic on the measured validation population, with all representation weights and statistics frozen.

Figure E.1: Recorded boxing by HoloBoxing (7-min version, red) against GGSD (top, match r02-m005) and RoboStriker (bottom, match r05-m015). The opponent is blue. Top frames occur at 3.02, 29.02, and 45.72 seconds; bottom frames at 3.02, 10.36, and 12.54 seconds. Hit labels follow recorded referee events; the visible poses are the corresponding control-frame states. Camera and lighting are adjusted; wall faces and corner posts are hidden for visibility.

Humanoid Robotics

Humanoid robotics develops human-shaped robots that move, manipulate objects, and interact with people in human environments.

See all 151 papers

Humanoid Hanoi: Investigating Shared Whole-Body Control for Skill-Based Box Rearrangement
Feb 14, 2026Minku Kim, Kuan-Chia Chen, Aayam Shrestha +3Humanoid Robot Control Whole-Body Control
The authors are with the Collaborative Robotics and Intelligent Systems Institute, Oregon State University, Corvallis, Oregon, 97331, USA
Learning Reusable Hybrid Motion Priors for Humanoid Locomotion from Motion Imitation
Jul 27, 2026Valerio Belli, Valerio Modugno, Enrico Mingo Hoffman +1RL for Robotics Latent Action Learning
UNIROMA, UCL · Sapienza University of Rome, Rome, Italy · University College London, London, UK +3
Continual Humanoid Motion Learning
Oct 3, 2026Zhewen He, Hao Huang, Geeta Chandra Raju Bethala +4Continual Robot Learning Humanoid Motion Tracking
New York University Abu Dhabi, Abu Dhabi, United Arab Emirates. · Fudan University, Shanghai, China.

Feb 14, 2026J/K move · Enter open · S save

Humanoid Hanoi: Investigating Shared Whole-Body Control for Skill-Based Box Rearrangement

Minku Kim, Kuan-Chia Chen, Aayam Shrestha, Li Fuxin+2

The authors are with the Collaborative Robotics and Intelligent Systems Institute, Oregon State University, Corvallis, Oregon, 97331, USA

Humanoid Robot Control Whole-Body Control Long-Horizon Robotic Manipulation Robot Manipulation Benchmarks

Open paper AskPDF arXiv

We introduce Humanoid Hanoi, a challenge problem for long-horizon humanoid skill composition that requires a robot to repeatedly sequence pickup, locomotion, and placement skills to rearrange stacked boxes in simulation and on hardware. This symbolically simple task exposes a weakness in a common modular humanoid architecture, where reusable high-level skills are learned on top of a shared, task-agnostic whole-body controller (WBC). While such a WBC provides a unified low-level interface for skill composition, it may not be sufficiently hardened to execute the full collection of learned skills under domain randomization and long-horizon closed-loop execution. We compare strategies for addressing this issue, including skill-specific WBC fine-tuning, residual low-level adaptation, and task-driven data aggregation that expands the shared-WBC training distribution with closed-loop trajectories induced by learned skills while preserving its original tracking objective. In our evaluated system, the shared-WBC approach achieves the strongest long-horizon robustness. A direct transition-smoothing intervention further shows that low-level controller handoffs account for a substantial part of the composition loss observed with non-shared controllers. Project page: https://osudrl.github.io/Humanoid_Hanoi/

Abstract

Humanoid Robotics

Explore similar work

Humanoid Hanoi: Investigating Shared Whole-Body Control for Skill-Based Box Rearrangement

Learning Reusable Hybrid Motion Priors for Humanoid Locomotion from Motion Imitation

Continual Humanoid Motion Learning