In a nutshell: Claude and other language models can handle simple robotics tasks when working through predefined controllers, but fail at direct motor control without additional abstraction.
Anthropic has investigated whether language models like Claude can control robots directly. Initial tests show: the capability depends heavily on how the model is connected to the robot.
Anthropic tested multiple language models on robot control, including a real four-legged Unitree Go2 robot system (used in Project Fetch) as well as simulations of a humanoid robot and a robot arm. Control methods varied in their abstraction: from direct motor torque control via Python controllers to reinforcement learning and guidance of predefined policies through language models.
The central finding is: robotics performance depends as much on the control function and robot type as on the model itself. When directly controlling motor torques, the models failed. However, when they supervised a pretrained controller or used simple navigation tools, they succeeded at real navigation and manipulation tasks. Tests included classic control tasks (pendulum balancing), locomotion (four-legged and humanoid robots walking) and manipulation (grasping and moving objects with a robot arm).
Progress between model generations is distributed unevenly: the most consistent improvements occur at more abstract interfaces, where predefined policies handle the underlying physics simulation. Current models can execute limited but meaningful full-body control tasks like balancing and walking legged robots, though with weaker progress on low-level humanoid robots.
Anthropic emphasizes the safety implication: today’s state-of-the-art systems cannot control humanoid robots without predefined policy, but newer models show tangible progress. A general chat model without specialized robotics training can perform practical tasks, such as guiding a four-legged robot through a maze or moving a plate from the counter to the stove – the reliability gap closes with each model generation.
Source: www.anthropic.com · Published 9 July 2026
Lumi AI News — AI-assisted curation according to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.