Robot Programs Complete 92% of Physical Trials

Researchers report AI-written robot programs succeeded in 221 of 240 physical evaluation trials across eight tasks.

Updated on Nov. 15, 2026 in Robotics

Isometric editorial illustration of a single robot arm hovering over a clean workbench beside a row of identical metal blocks, in oxblood, teal, and mustard.
Researchers reported that robot programs written by Anthropic's Claude Opus 5.5 coding agent completed 221 of 240 physical trials across eight tasks on November 12. AI Illustration. Upload story photo >

ACE-CaP-0 researchers reported that robot programs developed by Anthropic's Claude Opus 5.5 coding agent completed 221 of 240 physical evaluation trials, a 92.1% success rate. The study was posted November 12.

Why it matters

The results suggest AI coding agents can write physical robot control code that performs reliably when executed repeatedly without a coding agent or human in the loop. The findings bear on how far automated programming may extend into real-world robotics.

The physical setup used two Agilex PIPERX arms and three cameras, with charger insertion succeeding in all 30 trials and bread packing in 25. With simulation on demand, the charger task reached 100% success versus 86.7% without it, at a development time of 2.6 hours versus 3.7 hours.

The players

Anthropic

Anthropic is an artificial intelligence company that develops the Claude family of AI models, including the Claude Opus 5.5 coding agent used to write the robot programs.

ACE-CaP-0

ACE-CaP-0 is the research effort whose researchers posted the study evaluating AI-written robot programs on physical tasks.

The details

During development, the coding agent wrote, ran, inspected and revised its code using physical feedback, with human guidance and demonstrations available and at most ten human-assisted scene resets allowed per task. Each program was then frozen before its 30 separate evaluation trials.

Timeline

  1. November 12, 2026: ACE-CaP-0 researchers posted the study reporting the robot program results.

The Tech Race

The study extends the line of work, including Google DeepMind's RT-2 model, in which large AI models are applied to physical robot control. Here the AI's role shifts from directing the robot in real time to writing the program code that the robot later runs on its own.

For now the results are confined to a research setup of two arms, three cameras and eight tasks, so they do not yet change what consumers can buy or use. The next step the researchers describe is testing the frozen programs under new conditions to measure how broadly they can be reused.

The takeaway

Reliable AI-written robot code depends on physical testing with real feedback, not only on the code itself. Readers following automation should watch whether these programs hold up outside the controlled setup where they were built.

Further reading

Explore more on Robotics for the broader field of machines that sense and act in the physical world.