KAIST's RobotUse agent harness hits 45% task success on RoboLab, beating CaP-X
kaist-ai · hf · 2026-10-06
KAIST AI introduces RobotUse, a robot agent harness organizing computation, context, and decisions around specifying and revising physical actions:
- Agents visually select targets and poses while the backend handles geometry, motion planning, and control
- Subagents retain detailed interactions within each subgoal and return only what's needed for subsequent decisions
- Continual harnessing lets agents learn from execution by updating a persistent playbook
- 45% task success on RoboLab, outperforming CaP-X by 6.7 points, with compact decision contexts and less reliance on predefined action abstractions
- Demonstrates learning from real-world execution despite imperfect feedback, with transfer to later tasks
Project page: robotuse-team.github.io
More from Embodied
- OpenRUA: off-the-shelf Claude Code drives robots via ROS 2 at 99% success, zero training — Zhaoyang Chu · 2026-10-06
- AI is motion-capturing barbers' craft — which skilled trade gets recorded next? — Void_x7600 · 2026-10-06
- Noitom releases HiPHI: 617.5-hour human-object interaction dataset for humanoid robots — TinfoilTricorn · 2026-10-06
- September robotics funding hits $4.86B across 167 rounds, led by Cornerstone's $700M — lukas_m_ziegler · 2026-10-06
- Prompt-Driven Autonomous Drone Flies Fully Unsupervised After Start — AryHHAry · 2026-10-06
- Survey: 65% of Japanese seniors prefer robot-assisted nursing homes, willing to pay 8% more — HealthcareLdr · 2026-10-06