Humanoid boxing skills learned with seven minutes of gpu training
HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training
Robotics
Summary
Teaching a robot to box requires controlling its whole body precisely, which can be very complicated and time-consuming. This paper shows how to use an already trained body controller to simplify learning boxing moves by focusing on higher-level planning instead of all the joint movements. The authors demonstrate that by combining this approach with a small amount of additional training, the robot can learn to box effectively in just a few minutes on a single GPU. Their method separates general movement skills from specific boxing decisions, making the training faster and more efficient.
What this means in practice
- •For robotics developers: Train humanoid robots to perform complex tasks by efficiently building on pretrained whole-body controllers.
- •For animation software creators: Generate realistic humanoid motion sequences by separating low-level control from high-level action planning.
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.