Skip to main content
Robots now run AI-written code without an AI calling the shots, succeeding 92% of the time

Robots now run AI-written code without an AI calling the shots, succeeding 92% of the time

CryptoSlate
15 lis 2026, 14:46Cryptoslate
Oryginał
2 min read

Robots successfully completed 221 of 240 trials using Claude-written programs for tasks like uncapping bottles, with human help only during development, not execution.

ACE-CaP-0 researchers reported in a study published November 12 that robot programs developed by Anthropic's Claude Opus 5.5 coding agent completed 221 of 240 physical evaluation trials across eight different tasks. The results demonstrate that task-specific skills can persist in code and run without the coding agent choosing each action during execution.

The experiment separates program development from repeated execution. During development, the coding agent wrote, ran, inspected and revised its code using physical feedback, with the ability to ask for textual advice or demonstrations where a person physically guided the robot arms. However, during evaluation, the programs used live perception and control systems while their decision rules remained fixed.

Task Performance

Success varied across the eight tasks, with charger insertion succeeding in all 30 trials while bread packing, the weakest result, succeeded in 25. The physical setup used two Agilex PIPERX arms and three cameras to perform tasks including inserting a charger, uncapping a bottle and packing shoes. Mean physical execution times across tasks ranged from 0.97 to 2.27 minutes.

Development Process

During development, the agent could use simulated experiments when they helped resolve a development question. In a charger-insertion comparison, on-demand simulation produced a program with 100% evaluation success, versus 86.7% for development without simulation. Development took 2.6 hours with on-demand simulation versus 3.7 hours without it, while output-token usage increased from 730,000 to 1,090,000.

Development took 2.6 hours with on-demand simulation versus 3.7 hours without it, while output-token usage increased from 730,000 to 1,090,000.

ACE-CaP-0 researchers

Future Steps

The study provides evidence for skill reuse after human-assisted development, but establishing how consistently the development process produces equally capable programs requires separate development runs across tasks. Testing these routines under new conditions is the next step toward establishing how broadly they can be reused.

Z.ai
Article written by AI
Z.ai GLM-4.5-FlashFree
Free AI articles150 of 150
No limit after signing in with Telegram