Anthropic Research Shows Claude Language Models Can Control Real Robots
Key Takeaways
- ▸Language models can effectively control real robots when given appropriate abstraction levels and interfaces, not through direct motor control but through high-level policy supervision
- ▸Model capability in robotics depends as much on the robot body and control method as on the model itself—the same model performs differently depending on whether it's setting motor torques, writing controller code, or supervising a pretrained policy
- ▸Newer language models show consistent improvement on high-level robotics interfaces and measurable but uneven gains on direct low-level control tasks
Summary
Anthropic researchers, including Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman, published findings demonstrating that language models like Claude can effectively control robots across multiple embodiments, from simulated quadrupeds to real-world robotic arms and the Unitree Go2. The research tested models on classic control tasks, locomotion, navigation, and manipulation using varying levels of abstraction—from direct motor control to supervising pretrained policies.
A critical finding is that model performance depends heavily on the control interface used. When language models directly command motor torques, they largely fail. However, when supervising a pretrained policy or using higher-level abstraction tools, they successfully complete real navigation and manipulation tasks. Newer models show substantially improved ability to adjust strategies and convert visual and sensory understanding into appropriate robotic actions.
The research reveals uneven but consistent improvement across model generations, with the most reliable gains on high-level interfaces. Frontier models still cannot control humanoid robots without pretrained policies, but newer models have made meaningful progress in direct manipulation and high-level policy control. The authors note that general-purpose chat models with no robotics training can already perform tasks like navigating mazes or manipulating objects—capabilities that are improving with each generation.
- Real-world applications are emerging: models can navigate environments, manipulate objects, and perform complex tasks, though they still struggle with spatial memory and long open-loop planning
Editorial Opinion
This research marks a significant milestone in demonstrating that general-purpose language models can transfer their reasoning capabilities to physical robot control—a domain previously thought to require specialized training. The finding that abstraction level matters more than raw model capability suggests a pragmatic path forward for robotics deployment: pairing language models with pretrained controllers rather than attempting direct low-level control. As models continue to improve, the gap between AI systems operating in the digital and physical worlds continues to narrow, raising important questions about deployment safety and real-world integration that the authors rightfully flag as critical for the field.
