Coding agents hit 95% on robot planning

Claude Code and Codex reached mean success rates between 56% and 95% on generalized task-and-motion planning, beating hand-engineered planners’ 47% baseline across 16 comparable environments, in a study spanning 28 simulated environments and 98,000 evaluation episodes, according to the new TAMP paper. The result is a robotics signal, not a deployment guarantee: agents wrote programs that generalized to unseen instances, but the experiments stayed inside simulators and used a fixed synthesis budget. Operators tracking the archive’s physical-AI funding concentration should file this as software leverage while keeping hardware, calibration, and real-world failure testing separate.