SignalTech/AI
HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training
HoloBoxing enables humanoid boxing training in 7 minutes on a single GPU.
Context
What happened: HoloBoxing is a method for learning humanoid boxing that pairs a frozen whole-body controller with a compact planner predicting body-point targets, using imitation initialization and reinforcement learning.
Key numbers: About 7 minutes of additional single-GPU training, excluding controller pretraining, data preparation,
Why it matters: The recipe separates reusable motor skills from task decisions, so task optimization stays focused on a compact upper policy and avoids relearning whole-body control. Simulated boxing experiments and component controls evaluate the behavior.
What changed
HoloBoxing is a method for learning humanoid boxing that pairs a frozen whole-body controller with a compact planner predicting body-point targets, using imitation initialization and reinforcement learning.
Why it matters
The recipe separates reusable motor skills from task decisions, so task optimization stays focused on a compact upper policy and avoids relearning whole-body control. Simulated boxing experiments and component controls evaluate the behavior.