Claude Plays Robotics (www.anthropic.com)

Recent tests have explored the capabilities of language models, particularly the Claude series, in controlling various robotic systems, including classic control toys, a simulated quadruped, and a real Unitree Go2 robot. The research aimed to investigate whether the strengths of language models could translate into robotics, where precise 3D understanding and logical reasoning are essential. Various control methods were tested, ranging from direct motor torque commands to high-level steering instructions for pretrained robot policies. Significantly, while models often struggled with direct low-level control, they demonstrated better performance when interfacing with pretrained policies or higher-level commands, signaling that generational improvements in models such as Claude Opus 4.6 and Mythos Preview are narrowing the reliability gap in robotics. The implications of this research extend to the future development and safe deployment of language models in physical environments. Current frontier models can execute complex tasks like navigating a quadruped through an obstacle course with minimal prior training, suggesting a growing proficiency in both direct manipulation and high-level control across various robotic embodiments. Despite challenges, particularly with tasks requiring more degrees of freedom, the consistent performance improvements across generations signal a promising trajectory for integrating AI with robotics, potentially paving the way for more advanced applications in both industrial and home environments.