Event

HoloBoxing proposes a humanoid boxing learning recipe

2 reports · 2 sources

Xihan Ma and other authors submitted a preprint paper proposing the HoloBoxing humanoid robot boxing learning recipe: building on an existing pretrained whole-body controller, it separates reusable motor skills from task decision-making, with a compact planner outputting future body-point targets and the frozen controller responsible for joint-level execution. The authors state that, starting from this controller, the recipe requires about 7 minutes of additional training, including reusable representation fitting and single-GPU task runs; this time does not include controller pretraining, data preparation, evaluation, or search. A subsequent Reddit post claimed that the recipe can complete training from scratch in 7 minutes on a single consumer-grade GPU and released the code and paper, but that claim did not clarify, under the paper's conditions, whether "from scratch" includes controller pretraining. The relevant validation was carried out in simulated boxing experiments and component ablations, to evaluate the learned behaviors and the respective roles of imitation initialization and target representation.