prpl-group/agentic-gentamp-runs

AgenticGenTAMP agent runs

These are the runs behind Coding Agents for Generalized Task and Motion Planning Problems (arXiv:2609.30233). In each AgenticGenPlan run, a coding agent (Claude Code or Codex) gets a task description and access to a simulator. It writes a Python program (approach.py) within a fixed budget, committing to git as it goes. The final program is then evaluated on 100 held-out problem instances, including object counts larger than those the agent could test.

For each AgenticGenPlan run, the dataset includes:

• the full git history of the agent's sandbox, as a git bundle;
• the agent transcript and the experiment logs;
• the final program files;
• the held-out evaluation of the final program;
• a re-evaluation of every commit on the same held-out suite, which shows how performance changed during synthesis.

The Planner and LLMGenPlan baselines are included with their evaluations and logs.

Contents

method_id	method	model	runs	environments	mean solve rate
opus5	AgenticGenPlan	Opus 5 (high)	140	28	0.74
opus5-source	AgenticGenPlan + source	Opus 5 (high)	140	28	0.84
sol	AgenticGenPlan	GPT-5.6 Sol (medium)	140	28	0.51
astra	AgenticGenPlan	GPT-6 Astra (high)	140	28	0.86
astra-source	AgenticGenPlan + source	GPT-6 Astra (high)	140	28	0.95
opus55	AgenticGenPlan	Opus 5.5 (high)	140	28	0.95
opus55-source	AgenticGenPlan + source	Opus 5.5 (high)	140	28	0.96
llmgenplan	LLMGenPlan	Opus 5	140	28	0.28
planner	Planner		75	15	0.43

Environments come from KinDER and from PDDLStream domains (pr2blocked, pr2packed, rovers). Each method and environment pair has up to five runs, named by their run seed (24, 42, 222, 424 or 444).

Tables

All tables join on run_id, which has the form <method_id>/<environment>/seed<run seed>.

• runs: one row per run. Among the fields:
  • method_id, method, agent, model, reasoning_effort, environment and seed;
  • held-out results: solve_rate, by_count, design_count_solve_rate, held_out_count_solve_rate, num_crashed_episodes;
  • generation cost and usage: agent_cost_usd and the gen_* fields;
  • file information: num_commits, has_git_history, git_history_intact, has_transcript, and archive_path, which points to the run's raw files;
  • kinder_loophole and web_search_events, described under "Notes";
  • note, which explains the selection when a run differs from the default.
• episodes: the per-episode held-out results of each final program (solved, total_reward, num_steps, object_count, timing).
• commits: one row per commit in the agent's sandbox, with the subject, time, diff statistics and the patch (cut at 1 MB, see diff_truncated). The reeval_* fields give the commit's held-out solve rate.
• programs: the content of every tracked .py file at the final commit.
• checkpoints: one row per distinct program tree that was re-evaluated. Commits with the same tree share a row.
• checkpoint_episodes: the per-episode results of each re-evaluated tree.

from datasets import load_dataset

runs = load_dataset("prpl-group/agentic-gentamp-runs", "runs", split="train").to_pandas()
commits = load_dataset("prpl-group/agentic-gentamp-runs", "commits", split="train").to_pandas()

Raw run files

runs/<method_id>/<environment>/seed<N>.tar.gz holds the run folder:

• stream.jsonl: the agent's output stream;
• sandbox/: the agent's working tree, including approach.py, its probe scripts, and .agent_sessions/ for Codex session logs;
• results.json: the held-out evaluation;
• run_experiment.log and env_server.log;
• .hydra/: the experiment configuration;
• videos/, when present;
• sandbox.bundle: the full git history.

To restore the repository:

cd runs
tar xzf opus5/rovers/seed42.tar.gz
git clone opus5/rovers/seed42/sandbox.bundle repo
git -C repo log --stat

Evaluation

Every program is evaluated on the same 100 held-out instances with a 60 s per-episode policy timeout. The seeds that generate these instances are withheld so that the suite stays held out for future methods. The episodes are numbered 0 to 99 in the same order for every run.

Notes

• kinder_loophole, kinder_defects, loophole_strength and loophole_note mark runs where a manual audit found that the final program scores partly through a known defect in a KinDER success test. Runs that were not audited are not flagged.
• Three Sol ScoopPour3D runs used the model provider's web search. Seeds 24 and 222 were rerun without it, and seed 42 is kept with its searches counted in web_search_events. All three score 0.
• opus5/rovers/seed444 ships its raw .git folder, since one data object is truncated, and has no per-commit re-evaluation.
• sol/scooppour3d/seed222 used up its budget without submitting a program and scores 0.

License

The dataset is released under CC BY 4.0. Source code from KinDER, PDDLStream and other projects that appears in prompts, transcripts and sandboxes keeps its original license.

Citation

@article{merler2026codingagents,
  title   = {Coding Agents for Generalized Task and Motion Planning Problems},
  author  = {Merler, Matteo and Li, Bowen and Roy, Josh and Liang, Yichao and Wang, Qianwei and Huang, Yixuan and Silver, Tom},
  journal = {arXiv preprint arXiv:2609.30233},
  year    = {2026}
}