Robots Execute AI-Written Code Independently with 92% Success Rate

A new research study published on November 12 introduces ACE-CaP-0, a framework designed to develop robot control programs. The findings demonstrate that robot ai writing capabilities can produce specialized code that functions reliably without a coding agent managing the process in real time. In physical evaluation trials, these programs achieved a 92.1% success rate across eight distinct tasks, ranging from packing shoes to uncapping bottles.

Advancements in robot ai writing

The development process utilized Anthropic’s Claude Opus 5.5 coding agent, which wrote, inspected, and revised code by leveraging physical feedback. During this phase, the agent could request human demonstrations or textual advice to refine its logic. Human input included providing arm trajectories, video, and key joint poses to guide the robot's learning. The system also allowed for on-demand simulations to resolve development queries, which proved to enhance program performance compared to development conducted without simulation tools.

Once the development phase concluded, the resulting programs were frozen. In the subsequent evaluation phase, the robotic arms—specifically two Agilex PIPERX units—executed the tasks using internal perception and control systems. Crucially, the coding agent was removed from the execution loop. The programs operated autonomously based on fixed decision rules and real-time environmental observations. Out of 240 evaluation trials, the robots successfully completed 221, with charger insertion achieving a 100% success rate.

Robot programs developed by AI can now execute tasks independently with a 92.1% success rate.

Researchers noted a trade-off between development time and computational resources. Utilizing on-demand simulation reduced development time from 3.7 hours to 2.6 hours, though it increased output-token usage from 730,000 to 1,090,000. While these figures provide insight into the efficiency of the development cycle, the researchers emphasized that a comprehensive cost analysis must also account for the effort required to collect human demonstrations.

Why it matters

This research marks a shift in how robotic control programs are created and deployed. By separating the development phase from the execution phase, the ACE-CaP-0 method demonstrates that task-specific skills can be captured in code and reused repeatedly without requiring an active AI agent to oversee every movement. This approach provides a foundation for more scalable robotic automation.

The ability for robots to function with fixed decision rules after a human-assisted development phase suggests a pathway toward broader skill reuse. Future research will focus on testing these routines under new conditions to determine the versatility of the programs generated through this AI-driven development process.