Anthropic Tests Claude on Robotics: Direct Control Fails, LLM Supervision Hurts Familiar Tasks
DJiafei · x · 2026-07-30
Anthropic published a study on how LLMs like Claude perform on robotics tasks. They tested models controlling various embodiments—from classic control toys and a simulated quadruped to a robotic arm and a real Unitree Go2—using methods ranging from direct motor torque commands to writing controller code and providing high-level steering to pretrained policies.
Key Findings:
- Model capability depends heavily on how they are connected to the robot.
- Models mostly fail when they must drive the joints themselves.
- They can complete real navigation and manipulation tasks when supervising a pretrained controller or using simple orientation tools.
Interestingly, the study noted a counterintuitive finding highlighted by external testers: LLM supervisors can actually hurt performance on familiar tasks where a pretrained VLA model beats every LLM+VLA combo. However, for novel tasks the VLA can't solve alone, the best models provide a net uplift.
More from Embodied
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- Transforming quadruped robot runs on real hardware: per-pose RL policies in Isaac switched on morph — _Stocko_ · 2026-10-16
- Foldable iPhone Duo reportedly sells out launch-day inventory, some deliveries slip to 2027 — Polymarket · 2026-10-16
- KinoGen synthesizes robot loco-manipulation reference motions directly via kinematically grounded video generation — ChongZzZhang · 2026-10-16
- Science Robotics paper introduces F³, predictive event-based vision beyond optical flow — CSProfKGD · 2026-10-16
- iPhone Duo foldable appears on South African retailer iStore at R47,999 in apparent leak — shauntrennery · 2026-10-16