HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training
AuthorsXihan Ma, Bo Zhang, Boxiong Li, Yuhan Bao, Ziyue Liu, Yucheng Wang
A pretrained humanoid controller helps a compact planner learn simulated boxing in about seven minutes of additional single-GPU training.
Key results
Minutes, including reused target-representation fitting and the new task-training process.
Latent dimensions output by the feedback planner.
Future body-target coordinates produced by the decoder.
Final interaction-training updates in the short recipe.
Second place in the evaluated six-policy simulated boxing league.
Top ranking for the 21.4-minute training version.
What the paper found
HoloBoxing adapts a pretrained HoloMotion v2 whole-body controller for boxing on the Unitree G1, keeping the controller frozen while a compact feedback planner decides which movements to request. A learned decoder turns a 32-dimensional action into 225 future body-target coordinates; measured-state imitation first gives the planner an initial behavior, then PPO refines it through standing warmup and opponent interaction. The motion prior uses InterHuman and RoboStriker trajectories, and the complete recipe adds 128 PPO interaction updates. On a single NVIDIA RTX 4090, the reported adaptation takes 7.0 minutes, including reused target-representation fitting but excluding controller pretraining, data preparation, evaluation, and search. In a simulated six-policy boxing league, the short-training version ranks second with 1541 Elo; a 21.4-minute version ranks first with 1669 Elo. Component controls perform worse when imitation initialization or the latent decoder is removed, supporting the value of starting from measured-state behavior and retaining a structured target interface. The results demonstrate short task adaptation above reusable motor control, but remain simulation-only.
Original 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.
Read the original paperMore in Robotics
Browse all 59 papers →DiscoDemo: Discovering Efficient and Diverse Robot Demonstrations for Imitation Learning
Minho Park, Kinam Kim, Donghu Kim, Byungkun Lee, Dongyoon Hwang, Yongjae Shin, Junha Hyung, Hojoon Lee, Jaegul Choo
DiscoDemo trains robots to find varied successful ways to complete tasks, then turns those behaviors into better imitation-learning data.
HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second
Jonas Eschmann, Giuseppe Loianno
HyperDrone aims to make realistic drone simulation fast enough to train vision-based flight policies at massive scale, then transfer them to a real quadrotor.
Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts
Antonia Bronars, Younghyo Park, Kushal Kedia, Pulkit Agrawal
A robot learns to insert, mesh, and thread parts by feeling contact through its own joints, without vision or tactile sensors.