HoloBoxing
A method for training humanoid robot boxing strategies. It freezes a pretrained whole-body controller and trains only a lightweight planner, enabling rapid training on consumer-grade GPUs; it belongs to the hierarchical control paradigm.
Timeline (last 90 days)
The HoloBoxing project, shared on Reddit, provides a method for training a humanoid boxing strategy from scratch on a consumer-grade GPU.
HoloBoxing needs only 7 minutes to train a complete humanoid boxing strategy (on a single consumer-grade GPU).
HoloBoxing adopts a hierarchical control architecture: freezing a pretrained whole-body controller and training only a compact upper-layer planner.
Freezing the whole-body controller and training only the upper-layer planner can avoid the catastrophic forgetting common in end-to-end training.
HoloBoxing's 7-minute training time is achieved because gradients only pass through the lightweight planner, making backpropagation extremely cheap.
The HoloBoxing project has open-sourced both its code and its paper.
All Facts (6)
The HoloBoxing project, shared on Reddit, provides a method for training a humanoid boxing strategy from scratch on a consumer-grade GPU.
50%UnverifiedHoloBoxing needs only 7 minutes to train a complete humanoid boxing strategy (on a single consumer-grade GPU).
50%UnverifiedHoloBoxing adopts a hierarchical control architecture: freezing a pretrained whole-body controller and training only a compact upper-layer planner.
50%UnverifiedFreezing the whole-body controller and training only the upper-layer planner can avoid the catastrophic forgetting common in end-to-end training.
50%UnverifiedHoloBoxing's 7-minute training time is achieved because gradients only pass through the lightweight planner, making backpropagation extremely cheap.
50%UnverifiedThe HoloBoxing project has open-sourced both its code and its paper.
50%