HOST: Open-Source Robot Learns New Skills in 29s from Single Video, 507x Faster Than SFT
chris_j_paxton · x · 2026-08-03
X SQUARE ROBOT, in collaboration with Tsinghua University and BIT, has open-sourced HOST (Human-to-robot One-Shot Skill Acquisition). The framework enables robots to acquire new manipulation skills from a single human demonstration video during inference time.
Key Highlights:
- Rapid Acquisition: Learns from a video averaging just 29 seconds.
- Training-Free: Requires no task-specific fine-tuning or parameter updates, avoiding catastrophic forgetting.
- Mechanism: Instead of copying human trajectories, it aligns execution by task progress, predicts outcomes from the robot's perspective, and derives necessary actions to solve viewpoint and embodiment mismatches.
- Performance: Achieves a 62% success rate on 50 novel tasks, operates 507x faster than π0.5 + SFT, and retains 99% of prior performance.
Related event: HOST Framework Enables Robots to Learn Skills from 29-Second Videos(2 posts)→
More from Embodied
- ELLIS Institute Launches Physical AI Lab Focused on Robot Learning — orvieto_antonio · 2026-08-03
- Robot muscle material combines damage resistance and self-healing, researcher plans to scale up — MarwaEldiwiny · 2026-08-03
- Solving Humanoid Robotics: The Key Might Be Right Under Our Feet — MarwaEldiwiny · 2026-08-03
- Alibaba DAMO Academy Launches RynnBrain 1.1: Humanoid Robot Cooks Steak in One Take — CyberRobooo · 2026-08-03
- Pigey: Solving Long-Horizon Robotic Memory via Orchestrators Without Retraining — stepjamUK · 2026-08-03
- Is Folding Clothes the Holy Grail of Robotics? A New Heuristic Approach — DJiafei · 2026-08-03