Anthropic put Claude behind a real Unitree Go2 and a robot arm

Anthropic tested whether language model skills carry over to robotics by handing several models control of different bodies: classic control problems, a simulated quadruped and humanoid, a Franka Panda arm, and a real Unitree Go2 quadruped.

Results hinged on the level of abstraction. When a model set joint torques directly, it mostly failed. Working through a pretrained policy or simple orientation tools, models could clear real navigation and manipulation tasks.

Direct manipulation with the Franka Panda arm succeeded just 0 to 5.5% of the time. On the real Go2, every model failed an office-navigation task, and Grok 4.1 Fast spotted its target in the reflection of a glass door and drove straight at it. Anthropic says the code is coming to the safety-research/embody repo.