Anthropic Reveals Claude's Robotics Control Capabilities Experiment

로보틱스 | Tue Jul 14 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 1 sources

Anthropic's Frontier Red Team evaluated language models' robot control abilities across various embodiments and interfaces.

Analysis

[Anthropic Frontier Red Team] released Embody, a robotics control benchmark for Claude and language models ^[1]

• Tested classic control
• quadrupeds
• humanoids
• robot arms
• and real-world Unitree Go2
• Various abstraction levels from direct motor torque control to instructing pretrained policies
• Evaluated three domains: balancing
• locomotion
• and manipulation

[Language Model Robot Control Performance] identified performance gaps depending on control interface ^[1]

• Mostly failed when required to directly actuate joints
• Succeeded at real-world navigation and manipulation when supervising pretrained controllers or using orientation tools
• Newer models show improved ability to translate image and sensor understanding into actions

[AI Safety Implications] presented safety implications regarding frontier models' progress in physical control ^[1]

• Current frontier models cannot control humanoids without pretrained policies
• Even general-purpose chat models without robotics training can autonomously write tools to navigate a quadruped through a maze
• Reliability gaps between generations are rapidly narrowing

Sources

• [1] Frontier Red TeamClaude plays robotics - Anthropic Research