File size: 4,081 Bytes
bad7215 f8372ca bad7215 f8372ca 6df2ddf bad7215 f8372ca bad7215 f8372ca bad7215 f8372ca bad7215 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 | """Render the final five-task, four-method frozen-100-seed comparison table."""
from __future__ import annotations
import argparse
import hashlib
import json
from pathlib import Path
import matplotlib.pyplot as plt
TASKS = [
("LiftBarrier", 2, "lift_barrier"),
("CameraAlignment", 3, "camera_alignment"),
("ThreeRobotsStackCube", 3, "three_robots_stack_cube"),
("LongPipelineDelivery", 4, "long_pipeline_delivery"),
("TakePhoto", 4, "take_photo"),
]
METHODS = [
("冻结 DINOv3-ACT", "frozen_dinov3_act_all5_80k"),
("Stereo-ACT-cross_relbias", "stereo_cross_relbias_all5_80k"),
("Stereo-FFN-MoE", "stereo_ffn_moe_all5_80k"),
("Local-ARCA", "local_arca_all5_80k"),
]
def read_result(root: Path, method: str, task: str, seed_root: Path):
# The watcher writes one audited JSON per task beside the exact final
# checkpoint. Keeping the path tied to the training run prevents a stale
# result from a historical run from entering the formal All-5 table.
path = root / method / "formal_heldout_100" / f"eval_{task}.json"
if not path.is_file():
return None
raw = json.loads(path.read_text(encoding="utf-8"))
if raw.get("episodes") != 100:
raise ValueError(f"not formal 100-seed output: {path}")
expected = hashlib.sha256((seed_root / f"{task}.json").read_bytes()).hexdigest()
protocol = raw.get("seed_protocol", {})
if protocol.get("sha256") != expected or protocol.get("training_seed_overlap") != 0:
raise ValueError(f"seed audit failed: {path}")
return int(raw["successes"]), int(raw["episodes"])
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--root", default="/workspace/RoboFactory/runs/strict640x480_v2/results")
parser.add_argument("--output", default="/workspace/RoboFactory/runs/strict640x480_v2/final_report")
args = parser.parse_args(); root, out = Path(args.root), Path(args.output); out.mkdir(parents=True, exist_ok=True)
seed_root = root.parent / "heldout_seeds"
values = [[read_result(root, method_dir, task_id, seed_root) for _, _, task_id in TASKS] for _, method_dir in METHODS]
lines = ["| Training corpus | Test task (robots) | " + " | ".join(name for name, _ in METHODS) + " |",
"|---|---|" + "---|" * len(METHODS)]
table = []
for task_index, (task_name, robots, _) in enumerate(TASKS):
row = [column[task_index] for column in values]
best = max((item[0] / item[1] for item in row if item), default=None)
cells = []
for item in row:
if not item: cells.append("—"); continue
text = f"{item[0]}/{item[1]} ({100 * item[0] / item[1]:.1f}%)"
cells.append(f"**{text}**" if item[0] / item[1] == best else text)
lines.append("| All-5 strict640x480-v2 | " + f"{task_name} ({robots}) | " + " | ".join(cells) + " |")
table.append(cells)
(out / "performance_table.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
(out / "formal_seed_audit.json").write_text(json.dumps({
"status": "PASS", "episodes_per_cell": 100,
"tasks": [task_id for _, _, task_id in TASKS],
"methods": [method_id for _, method_id in METHODS],
"condition": "each result uses the matching frozen manifest with zero training-seed overlap",
}, indent=2) + "\n", encoding="utf-8")
fig, axis = plt.subplots(figsize=(16, 4.8)); axis.axis("off")
rendered = axis.table(cellText=table, colLabels=[name for name, _ in METHODS],
rowLabels=[f"{name} ({robots})" for name, robots, _ in TASKS],
cellLoc="center", loc="center")
rendered.auto_set_font_size(False); rendered.set_fontsize(10); rendered.scale(1, 2.0)
axis.set_title("RoboFactory — strict wrist-only RGB-D, All-5, frozen unseen 100-seed evaluation", pad=20, fontsize=14, weight="bold")
fig.tight_layout(); fig.savefig(out / "performance_table.png", dpi=220, bbox_inches="tight")
print(out / "performance_table.png")
if __name__ == "__main__":
main()
|