HoloBoxing trains humanoid boxing in about seven minutes on one GPU
Original title: HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training
A frozen whole-body controller executes joints while a compact planner outputs future body-point targets via a 32-dimensional decoder.
HoloBoxing is a training recipe that adds humanoid boxing on top of a pretrained whole-body controller. A compact planner requests future body-point targets, a learned decoder maps 32-dimensional actions to target windows, and measured-state imitation initializes the planner before reinforcement learning. The frozen controller and reused decoder avoid relearning whole-body control. The authors say the added training, excluding controller pretraining and several other stages, is about seven minutes on one GPU, evaluated in simulation.
Unconfirmed: The seven-minute claim is the authors' accounting and excludes controller pretraining.
- Task learning reuses the decoder and controller instead of relearning whole-body control.
- The seven-minute figure excludes pretraining, data preparation, evaluation, and search.
- Imitation initialization supplies behavior before reward-driven exploration.
- Experiments are simulated boxing runs plus component controls.
From the original article
1 min read at huggingface.co
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.
By Xihan Ma, Bo Zhang, Boxiong Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.