HoloBoxing trains humanoid boxer in 7 minutes on one GPU

Teaching a simulated humanoid to box takes about seven minutes of additional single-GPU training, HoloBoxing claims, provided a whole-body controller has already been trained separately. The arXiv preprint, posted October 23, 2026 by Xihan Ma and five co-authors, is explicit that the figure "excludes controller pretraining, data preparation, evaluation, and search."

The recipe freezes an existing whole-body controller and layers what the authors call a "compact planner" on top. The planner "requests future body-point targets," and a learned decoder "maps 32-dimensional actions to structured target windows." Before reinforcement learning begins, "measured-state imitation initializes the planner." Task-specific RL then shapes "interaction behavior while reusing the decoder and controller, avoiding the need to relearn whole-body control."

The experiments described are simulated, with component ablations testing the roles of imitation initialization and the target representation. No real-robot results, hit-rate numbers, or baseline comparisons appear in the abstract. The authors position the work as "a simple recipe," not a performance claim.