HoloBoxing teaches humanoid boxing with just 7 minutes of single-GPU training
Xihan Ma · hf · 2026-10-26
- Horizon Robotics' HoloBoxing is a simple recipe for humanoid boxing that separates reusable motor skills from task decisions: a frozen whole-body controller handles joint-level execution while a compact planner predicts body-point targets.
- A learned decoder maps 32-dimensional actions to structured target windows; measured-state imitation initializes the planner, then RL refines interaction behavior while reusing the decoder and controller.
- Excluding controller pretraining, data prep, and evaluation, task learning takes only 7 minutes on a single GPU. Code open-sourced in the HoloMotion repo.
More from Embodied
- HKU/ZJU/Fudan with Lightspeed launch HiSTEVE: hierarchical VLA agent playing Minecraft via native mouse and keyboard — tencent · 2026-10-26
- Hobbyist teaches StackChan robot to read scripts with Grok and Hermes-Agent — alexcovo_eth · 2026-10-26
- 10 wild humanoid robots: ex-Boston Dynamics droid butler, car-lifting bodyguard bot, ladder-climbing G1 — minchoi · 2026-10-26
- KAIST's humanoid v0.9 pushes 35 kg and lifts 20 kg, then moonwalks — minchoi · 2026-10-26
- Echo teaches VLA robots to detect task success and failure without frame-level labels — Yufeng Liu · 2026-10-26
- Alibaba DAMO releases 16B unified embodied model RynnUnify, 51.5% on RoboCasa365 — Minghao Zhu · 2026-10-26