""" Batch runner for the HandDrag trajectory-tracking baseline. Run this from an activated isaacgym environment: python scripts/run_trajectory_baselines.py --phase drag Each object gets: output/baselines/trajectory_tracking/.csv output/baselines/trajectory_tracking/.json output/baselines/trajectory_tracking/.log The aggregate table is written to: output/baselines/trajectory_tracking/summary.csv """ import argparse import csv import json import os import subprocess import sys from pathlib import Path import yaml PROJECT_ROOT = Path(__file__).resolve().parents[1] DEFAULT_OUT_DIR = PROJECT_ROOT / "output" / "baselines" / "trajectory_tracking" DEFAULT_MANIFEST = PROJECT_ROOT / "data" / "manifest.csv" def parse_args(): parser = argparse.ArgumentParser( description="Run trajectory-tracking baselines over saved HandDrag objects" ) parser.add_argument("--config", default="configs/env/hand_config.yaml") parser.add_argument("--manifest", default=str(DEFAULT_MANIFEST)) parser.add_argument("--dataset_root", default=None, help="Legacy layout root containing /trajectory.json. If omitted, use --manifest.") parser.add_argument("--out_dir", default=str(DEFAULT_OUT_DIR)) parser.add_argument("--object_ids", nargs="*", default=None) parser.add_argument("--phase", default="drag") parser.add_argument("--max_frames", type=int, default=None) parser.add_argument("--sim_steps_per_frame", type=int, default=2) parser.add_argument("--control_mode", choices=("pd_target", "delta_action"), default="pd_target") parser.add_argument("--settle_steps", type=int, default=8) parser.add_argument("--hold_final_steps", type=int, default=0) parser.add_argument("--break_on_done", action="store_true") parser.add_argument("--no_skip_missing_assets", action="store_true", help="Fail on missing GAPartNet assets instead of recording missing_asset") return parser.parse_args() def load_manifest_samples(manifest_path: Path, object_ids=None): selected = set(str(x) for x in object_ids) if object_ids else None samples = [] with open(manifest_path, newline="") as f: for row in csv.DictReader(f): if row.get("enabled", "1").strip() != "1": continue object_id = row["object_id"] if selected and object_id not in selected and row.get("sample_id") not in selected: continue samples.append({ "sample_id": row.get("sample_id") or object_id, "object_id": object_id, "trajectory": Path(row["trajectory"]), }) return samples def discover_object_ids(dataset_root: Path): return sorted( p.parent.name for p in dataset_root.glob("*/trajectory.json") if p.parent.name.isdigit() ) def load_config(config_path: Path): with open(config_path) as f: return yaml.safe_load(f) def asset_path(cfg, object_id: str): asset_cfg = cfg["asset"] return ( Path(asset_cfg["asset_root"]) / asset_cfg["arti_obj_root"] / object_id / "mobility_annotation_gapartnet.urdf" ) def baseline_env(): env = os.environ.copy() isaacgym_root = os.environ.get("ISAACGYM_ROOT") isaacgym_python = Path(isaacgym_root) / "python" if isaacgym_root else None if isaacgym_python and isaacgym_python.exists(): old_pythonpath = env.get("PYTHONPATH") env["PYTHONPATH"] = ( str(isaacgym_python) if not old_pythonpath else f"{isaacgym_python}:{old_pythonpath}" ) conda_prefix = env.get("CONDA_PREFIX") if conda_prefix: lib_dir = str(Path(conda_prefix) / "lib") old_ld = env.get("LD_LIBRARY_PATH") env["LD_LIBRARY_PATH"] = lib_dir if not old_ld else f"{lib_dir}:{old_ld}" py_tag = f"py{sys.version_info.major}{sys.version_info.minor}" env.setdefault("TORCH_EXTENSIONS_DIR", f"/tmp/torch_extensions/{py_tag}_cu121") return env def make_command(args, object_id: str, trajectory: Path, metrics_csv: Path, summary_json: Path): cmd = [ sys.executable, str(PROJECT_ROOT / "scripts" / "track_trajectory_baseline.py"), "--config", args.config, "--object_id", object_id, "--trajectory", str(trajectory), "--phase", args.phase, "--sim_steps_per_frame", str(args.sim_steps_per_frame), "--control_mode", args.control_mode, "--settle_steps", str(args.settle_steps), "--hold_final_steps", str(args.hold_final_steps), "--log_csv", str(metrics_csv), "--summary_json", str(summary_json), ] if args.max_frames is not None: cmd.extend(["--max_frames", str(args.max_frames)]) if args.break_on_done: cmd.append("--break_on_done") return cmd def tail(text: str, n_lines: int = 20): return "\n".join(text.splitlines()[-n_lines:]) def main(): args = parse_args() out_dir = Path(args.out_dir) out_dir.mkdir(parents=True, exist_ok=True) cfg = load_config(PROJECT_ROOT / args.config) if args.dataset_root: dataset_root = Path(args.dataset_root) samples = [ {"sample_id": object_id, "object_id": object_id, "trajectory": dataset_root / object_id / "trajectory.json"} for object_id in (args.object_ids or discover_object_ids(dataset_root)) ] else: samples = load_manifest_samples(Path(args.manifest), args.object_ids) env = baseline_env() rows = [] for sample in samples: object_id = sample["object_id"] sample_id = sample["sample_id"] trajectory = sample["trajectory"] metrics_csv = out_dir / f"{sample_id}.csv" summary_json = out_dir / f"{sample_id}.json" log_path = out_dir / f"{sample_id}.log" base_row = {"sample_id": sample_id, "object_id": object_id, "status": "pending", "error": ""} if not trajectory.exists(): base_row.update(status="missing_trajectory", error=str(trajectory)) rows.append(base_row) continue asset = asset_path(cfg, object_id) if not asset.exists() and not args.no_skip_missing_assets: base_row.update(status="missing_asset", error=str(asset)) rows.append(base_row) continue cmd = make_command(args, object_id, trajectory, metrics_csv, summary_json) proc = subprocess.run( cmd, cwd=PROJECT_ROOT, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, ) log_path.write_text(proc.stdout) if proc.returncode != 0: base_row.update(status="failed", error=tail(proc.stdout)) rows.append(base_row) continue with open(summary_json) as f: summary = json.load(f) summary["status"] = "ok" summary["error"] = "" rows.append(summary) fieldnames = [ "sample_id", "object_id", "status", "phase", "target_joint_idx", "handle_link_name", "start_frame", "end_frame", "sim_steps", "final_target_joint", "expert_target_joint", "selected_expert_delta", "normalized_progress", "mean_palm_handle_dist", "success_step_fraction", "error", ] summary_csv = out_dir / "summary.csv" with open(summary_csv, "w", newline="") as f: writer = csv.DictWriter(f, fieldnames=fieldnames, extrasaction="ignore") writer.writeheader() writer.writerows(rows) ok = sum(row.get("status") == "ok" for row in rows) print(f"wrote {summary_csv}") print(f"completed {ok}/{len(rows)} objects") if __name__ == "__main__": main()