HoloBoxing: Learn humanoid boxing in seven minutes of single GPU training

By reusing an already trained whole body controller, a team limits boxing learning to a small planning policy, trained in minutes.

Rather than relearn all humanoid control for each new task, HoloBoxing separates reusable motor skills from task-specific decisions. A pre-entry whole body controller, frozen, manages joint execution by joint. Above, a compact planner simply requests future positions for certain points of the body, via a learned decoder that translates 32-dimensional actions into structured target windows.

The planner is initially initialized by imitation of measured movements, which gives it a starting behavior, then refined by learning by strengthening to manage the interaction proper to boxing. Once the basic controller is available, the set requires only about seven minutes of additional training on a single GPU, including decoder adjustment. The authors state that this figure excludes controller pre-training, data preparation and evaluation.

The simulation experiments analyse the behaviour obtained and the respective role of imitation initialization and target representation.