NTU team has Claude write robot control code, hitting 70.24% success in sim
aigclink · x · 2026-10-14
A team from Nanyang Technological University and collaborators proposes COAP under an "embodied Turing machine" framing: a coding agent powered by Claude Opus 5.5 develops robot control programs offline, with no VLM, VLA, or online agent decisions at execution time.
- How it differs: VLA models map observations directly to actions and Agent-as-Policy approaches reason online; COAP moves the LLM out of the real-time control loop and uses it only for offline policy development and improvement.
- Execution: code continuously updates environment and task state from cameras and proprioceptive sensors, decides actions via explicit rules, and supports failure recovery rather than playing back fixed trajectories.
- Results: 70.24% success across 42 bimanual tasks on RoboDojo, up 38.9 points from the previous best of 31.38%.
- Caveats: all results are simulation-only, relying on known object geometry and camera setups; real-world contact complexity, lighting changes, unknown objects, and generalization remain open challenges.
More from Embodied
- Nature: smart glasses are heading for a major privacy crisis — marigo · 2026-10-14
- Impatient for JetKVM Mini, dev buys LuckFox KVM to let agents control his Mac Mini — TheZachMueller · 2026-10-14
- China's electric autonomous tractor T70 goes viral, already deployed in Hebei — lukas_m_ziegler · 2026-10-14
- Nomagic teases more updates as warehouse robots tackle unexpected scenarios — m_wulfmeier · 2026-10-14
- Rabbit R1 called 'ahead of its time' as modders plan to hack the once-mocked AI device — SimonBalmain · 2026-10-14
- RSI-Nav: recursive self-improvement for navigation agents hits 68.6% SR on R2R-CE — OpenGVLab · 2026-10-14