HoloBoxing: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training

HoloBoxing. A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training. By Xihan Ma and five others. Published on the twenty-third of October. Its code repository is worked out: about six hundred source files, with a readme and a licence.

Teaching a humanoid to box involves two different kinds of competence: control of the whole body, and the task decisions of the bout itself. A task learner that has to acquire both at once has to relearn whole-body control along the way, even when a controller that already handles it is available.

HoloBoxing is a simple recipe that builds on a pretrained whole-body controller and keeps reusable motor skills separate from task decisions. A compact planner requests future body-point targets. The controller, kept frozen, handles execution at the level of the joints. The planner decides where the body should go, and the controller decides how the joints get it there.

Between the two sits a learned decoder, which maps thirty-two-dimensional actions to structured windows of targets. Before any reinforcement learning, the planner is initialized by imitation of measured states. Task learning then refines the interaction behavior, reusing the decoder and the controller as they are. Task optimization stays focused on a compact upper policy, and reward-driven exploration starts from a behavior that already exists.

Starting from that available controller, the recipe accounts for about seven minutes of additional training. That figure includes reused representation fitting and a single-GPU task run. It excludes controller pretraining, data preparation, evaluation and search. The resulting behavior is evaluated in simulated boxing experiments, alongside component controls that test the roles of imitation initialization and of the target representation.

Seven minutes on a single GPU is the number in the title, and it covers the extra training stacked on a controller that already exists, not the controller itself. That boundary is the design. Whole-body skill is paid for once, in pretraining, and boxing is learned as a compact layer of decisions above it.