| """Check chunk_eval.py against lerobot-eval: same checkpoint, same seeds, zero delay, no displacement. |
| |
| For each execution mode, lerobot-eval runs ACT with the matching n_action_steps / |
| temporal_ensemble_coeff override, while chunk_eval.py runs the same mode through ChunkExecutor. |
| Matching per-seed outcomes and summed rewards show that the harness adds no behavior of its own. |
| |
| Usage: |
| python scripts/compare_harness.py --lerobot OUT/lerobot --harness OUT/harness --out summary.json |
| """ |
|
|
| import argparse |
| import json |
| from pathlib import Path |
|
|
| PAIRS = {"full": "full_d00_m0", "replan25": "replan25_d00_m0", "replan10": "replan10_d00_m0", "te": "te_d00_m0"} |
|
|
|
|
| def lerobot_episodes(eval_info: Path, start_seed: int) -> dict[int, tuple[bool, float]]: |
| metrics = json.loads(eval_info.read_text())["per_task"][0]["metrics"] |
| pairs = zip(metrics["successes"], metrics["sum_rewards"], strict=True) |
| return {start_seed + i: (bool(s), float(r)) for i, (s, r) in enumerate(pairs)} |
|
|
|
|
| def harness_episodes(condition_dir: Path) -> dict[int, tuple[bool, float]]: |
| episodes = {} |
| for path in sorted(condition_dir.glob("seeds_*.json")): |
| for e in json.loads(path.read_text())["episodes"]: |
| episodes[e["seed"]] = (bool(e["success"]), float(e["sum_reward"])) |
| return episodes |
|
|
|
|
| def main() -> None: |
| parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) |
| parser.add_argument("--lerobot", type=Path, required=True, help="folder with <mode>/eval_info.json") |
| parser.add_argument("--harness", type=Path, required=True, help="chunk_eval.py --out folder") |
| parser.add_argument("--start-seed", type=int, default=1000) |
| parser.add_argument("--out", type=Path, default=None) |
| args = parser.parse_args() |
|
|
| rows = [] |
| for mode, condition in PAIRS.items(): |
| ref = lerobot_episodes(args.lerobot / mode / "eval_info.json", args.start_seed) |
| ours = harness_episodes(args.harness / condition) |
| seeds = sorted(ref.keys() & ours.keys()) |
| row = { |
| "mode": mode, |
| "n": len(seeds), |
| "lerobot_successes": sum(ref[s][0] for s in seeds), |
| "harness_successes": sum(ours[s][0] for s in seeds), |
| "same_outcome": sum(ref[s][0] == ours[s][0] for s in seeds), |
| "same_sum_reward": sum(abs(ref[s][1] - ours[s][1]) < 1e-6 for s in seeds), |
| } |
| rows.append(row) |
| print(row, flush=True) |
| if args.out: |
| args.out.write_text(json.dumps(rows, indent=1)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
