Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence
Researchers propose Physical Coding so robot agents store task state as revisable code and beat XR-1 on RoboCasa365.
Vision-language-action policies encode task state inside action sequences, so small layout or viewpoint changes cause failure. The authors propose Physical Coding: Code as World records objects, relations, constraints, and progress, while Code as Policy handles planning, verification, recovery, and execution. HexaAnything calls perception, planning, control, and VLA/WAM tools and stores verified traces as memory. On RoboCasa365 it improves Composite-Unseen and overall success versus XR-1, and a harness-trained HexaModel beats its base on every split; it also completes PhyBench experiments and most tasks on a dual-arm AgileX robot.