HoloBoxing Separates Humanoid Boxing Planning from Whole-Body Control

Humanoid boxing is difficult because the robot must coordinate more than its arms. A punch depends on torso motion, leg support, balance, timing, and responses to an interaction that can change over time. If reinforcement learning starts directly from joint commands, the policy must discover both basic whole-body coordination and boxing behavior, making exploration unnecessarily expensive.

HoloBoxing proposes a separation between these responsibilities. The method assumes a pretrained whole-body controller and keeps it frozen during task learning. This controller translates higher-level body targets into joint-level movement, so the task policy does not need to rediscover fundamental coordination. A compact planner instead focuses on deciding where relevant body points should move in the near future.

The action interface is deliberately structured. A learned decoder maps a 32-dimensional action into target windows for body points, rather than exposing every joint as a direct decision variable. This gives the planner a smaller and more organized space in which to make task decisions. It also lets the lower controller handle the detailed execution and coordination required to realize those targets.

Training is staged. First, measured-state imitation is used to initialize the planner. The purpose is to provide a usable behavior before reward-driven exploration begins, rather than asking reinforcement learning to discover everything from scratch. Reinforcement learning then refines the interaction behavior for the boxing task. The decoder and whole-body controller are reused, keeping optimization concentrated on the compact upper-level policy.

The paper reports roughly seven minutes of additional training on a single GPU once the relevant whole-body controller is available. That figure includes reused representation fitting and the task run, but excludes controller pretraining, data preparation, evaluation, and search. It therefore should not be read as the total cost of creating a boxing humanoid from zero. The reported experiments are simulated, and component studies examine the contributions of imitation initialization and the target representation.

The broader lesson is an engineering one: task learning may become substantially simpler when reusable motor competence is separated from task-specific planning. A similar design could be useful for other humanoid behaviors that require coordinated movement and interaction, provided the underlying controller already supports the necessary motions. At the same time, the approach does not by itself establish general-purpose humanoid control. Its transfer depends on the coverage of the pretrained controller, the quality of the target representation, and the gap between simulation and physical robots.

HoloBoxing is therefore best understood as a compact training recipe, not a claim that humanoid boxing has become a solved problem. Its contribution is to show how a frozen execution layer and an initialized planner can narrow the learning problem to task behavior.

Source: Hugging Face Daily Papers