Coding agents synthesize programs for simulated TAMP tasks
Three agent configurations achieved mean success of 56–95% in the comparable simulation evaluation; this is not one model’s easiest-to-hardest task range.
Coding Agents for Generalized Task and Motion Planning Problems evaluates 980 generated programs on 100 held-out instances each, totaling 98,000 episodes across 28 simulated environments. Agents interact with a simulator while synthesizing a program within a fixed budget; the frozen program is then evaluated on unseen instances.
The reported mean success range of 56–95% belongs to three agent configurations: Claude Code with Opus 5 and Codex with GPT-5.6 Sol or GPT-6 Astra. The hand-engineered planner comparison is 47% on the 16 environments where a planner is available. The original article misassigned 95% and 56% to one model’s easiest and hardest suites.
The authors report lower average computation per evaluated instance as object counts grow. Synthesis expense remains a separate budget, and simulation results do not demonstrate reliability on a physical factory floor.
The companion GRASP work uses macro-generation, local revision and verification. Its reported 12.4 and 30.8 improvements are accuracy gains, not dollar savings; the original article inserted dollar signs and incorrectly described the macro stage as a traditional planner. A fair comparison must retain each work’s tasks, metrics and computational budgets.
Editorial correction (2026-10-13 UTC): Corrected configuration averages and GRASP accuracy units.