chess-sim/code/sim/push_dataset.py

"""Upload phase 2 outputs into one folder of a Hugging Face dataset repo.

Everything goes under `--path-in-repo` (default `chess-sim`) and nothing outside
that folder is touched:

  chess-sim/README.md             what is here and how to load it
  chess-sim/datasets/<name>/      the LeRobot dataset with phase2_episodes.jsonl and
                                  phase2_summary.json (LeRobot's temporary images left out)
  chess-sim/review/<name>.mp4     the review video, if the dataset folder has one
  chess-sim/reports/              sim/reports/ (grasp validation)
  chess-sim/code/                 the sim/ code and assets that made the data, and pyproject.toml

LeRobot's own push_to_hub writes to the repo root and replaces its README, so the
folder is staged locally and sent with upload_large_folder, which only adds or
replaces files. Log in once first with `.venv/bin/hf auth login` (a token that can
write to the repo).

Run:  .venv/bin/python sim/push_dataset.py --repo-id Machanize/playful --name pilot_60
"""
from __future__ import annotations

import argparse
import json
import os
import shutil
import sys
import tempfile
from pathlib import Path

HERE = Path(__file__).resolve().parent
ROOT = HERE.parent
sys.path.insert(0, str(HERE))

SKIP_DIRS = {"__pycache__", "renders", "reports", ".cache"}


def _place(src: Path, dst: Path):
    dst.parent.mkdir(parents=True, exist_ok=True)
    try:
        os.link(src, dst)            # same disk: no copy of a large dataset
    except OSError:
        shutil.copy2(src, dst)


def _tree(src: Path, dst: Path, skip=frozenset()):
    for dirpath, dirnames, filenames in os.walk(src):
        dirnames[:] = [n for n in dirnames if n not in skip]
        for f in filenames:
            if f == ".DS_Store":
                continue
            p = Path(dirpath) / f
            _place(p, dst / p.relative_to(src))


NOTES = {
    "pilot_60": "The first pilot: nominal board and tray, narrower lighting and camera variety. Kept for reference.",
    "varied_2000": "Heavy randomisation (see *What varies*). Trained the first model, 0% in the sim test "
                   "(since deleted; report in `reports/policy_eval/`).",
    "baseline_v1": "The simple baseline (see *Baseline scene*): one fixed scene and ten fixed opening moves. "
                   "Trains `Machanize/chess_phase_smolvla_baseline`.",
    "baseline_recovery_v1": "Correction demonstrations in the baseline scene (see *Corrections*), first "
                            "attempt: the teacher rises and pauses before regrasping, which made the fine-tuned "
                            "policy hover (20%). Kept for reference; not used for training.",
    "baseline_recovery_v2": "Correction demonstrations, second attempt (`collect_recovery.py --style v2`): the "
                            "teacher lines up from where the arm is, rising only as far as needed, and goes "
                            "straight down, with no pause. Fine-tuning the baseline on these plus baseline_v1 "
                            "lowered success (30% vs 51.7% on the same episodes: more pieces knocked over). "
                            "Kept for reference; not used.",
    "baseline_dart_v1": "baseline_v1's 300 scenes (same episode seeds) with a perturbed teacher (see *Perturbed "
                        "teacher*): in about half the episodes the executed approach drifts 2-10 mm off-centre and "
                        "returns before the fingertips reach the piece, while the recorded actions stay the clean "
                        "plan. A model trained fresh on it with the baseline recipe corrects on the way down and "
                        "closes more precisely (5.4 vs 6.6 mm); success 28.5% vs 25.5% on 200 paired scenes "
                        "(not significant), 17/40 vs 7/40 on the trace scenes (`reports/dart_eval/`). Checkpoint in "
                        "`experiments/dart/`.",
}


def _info(repo_id: str, folder: str, name: str) -> dict:
    """LeRobot's meta/info.json of a dataset already in the repo."""
    from huggingface_hub import hf_hub_download

    try:
        return json.loads(Path(hf_hub_download(repo_id, f"{folder}/datasets/{name}/meta/info.json",
                                               repo_type="dataset")).read_text())
    except Exception:
        return {}


def readme(repo_id: str, folder: str, name: str, dataset: Path) -> str:
    """The folder README: every dataset in it (those already uploaded plus this one)."""
    from huggingface_hub import HfApi

    try:
        names = {Path(e.path).name for e in HfApi().list_repo_tree(repo_id, path_in_repo=f"{folder}/datasets",
                                                                   repo_type="dataset")}
    except Exception:
        names = set()
    names.add(name)
    val_path = HERE / "reports" / "grasp_validation.json"
    val = json.loads(val_path.read_text())["summary"] if val_path.exists() else {}
    lines = [
        "# chess-sim: SO-101 chess pick-and-place in simulation", "",
        "Scripted-expert demonstrations of an SO-101 arm moving one chess piece from a square "
        "to another square or into a capture tray, recorded in MuJoCo. Made for training SmolVLA "
        "(phase 2 of the play-machanize project).", "",
        "Every dataset is LeRobot v3.0 at 30 fps, 640x480 images, with the source square highlighted "
        "red and the destination (square or tray) blue. `observation.state` and `action`: 5 arm joints in "
        "degrees and the gripper 0-100 (LeRobot's SO-101 follower with `use_degrees=True`). One instruction "
        "for every frame: \"move the piece on the red square to the blue square\". `phase2_episodes.jsonl` "
        "has per-episode notes (seed, piece, squares, board, piece set, lighting, cameras).", "",
        "## Datasets", "",
        "| folder | episodes | frames | cameras | video | notes |", "|---|---|---|---|---|---|",
    ]
    for n in sorted(names):
        info = json.loads((dataset / "meta" / "info.json").read_text()) if n == name else _info(repo_id, folder, n)
        cams = [k.split(".")[-1] for k, f in info.get("features", {}).items() if f.get("dtype") == "video"]
        vid = next((f.get("info", {}) for f in info.get("features", {}).values() if f.get("dtype") == "video"), {})
        video = f"{vid.get('video.codec', '?')} {vid.get('video.pix_fmt', '')}".strip()
        lines.append(f"| `datasets/{n}/` | {info.get('total_episodes', '?')} | {info.get('total_frames', '?')} | "
                     f"{', '.join(cams) or '?'} | {video} | {NOTES.get(n, '')} |")
    lines += ["", "`review/<name>.mp4`: random episodes of a dataset, read back through LeRobot. "
              "`reports/`: expert validation, policy tests and diagnostics. `code/`: the simulation, expert "
              "and export code that produced the data.", ""]
    lines += ["## What varies (varied_2000)", "",
              "- **Board and tray:** 22-25 mm squares, 6-22 mm border, 3-20 mm thick, pieces scaled to the "
              "squares. The robot plays white or black. The board's angle and position vary within reach; "
              "the tray varies in side, angle, size and colour.",
              "- **Table:** slides so its edges and corners, with the floor, come into view. Matte to glossy "
              "tables, with varied floors and walls.",
              "- **Lighting:** one to four lamps (window, ceiling, desk lamp, sun), 2700-6500 K, dim to "
              "bright, with and without shadows.",
              "- **Colours:** the arm (white most often), the pieces and the board.",
              "- **Clutter:** a laptop, phone, mug, notebook, pen and cables around the board.",
              "- **Overhead camera:** anywhere a person would mount it (35-80 cm up, up to 45 degrees from "
              "vertical, any side but behind the robot, 38-62 degree lenses). The wrist and rover cameras "
              "vary only by small mounting errors.",
              "- **Webcam look:** auto-exposure, white balance, blur, noise, JPEG and vignetting.",
              "- **Video:** LeRobot's default encoding (H.264, CRF 30, 4:2:0 chroma), which washes out the "
              "small red and blue squares.", ""]
    if "baseline_v1" in names:
        lines += ["## Baseline scene (baseline_v1)", "",
                  "Everything fixed, to get reliable closed-loop success before adding variety back one change "
                  "at a time (settings: `code/sim/phase2_baseline.toml`).", "",
                  "- The nominal board (25 mm squares) square to the robot and flush against its deck; the "
                  "robot plays black. Reference piece set, black and white pieces, white arm.",
                  "- Tray on the right, fixed table, the scene's window and ceiling lamps, no clutter.",
                  "- An overhead camera like the owner's (65 cm up, straight down, 44 degree lens, the arm at the "
                  "top edge) and the wrist camera, no mounting error, a clean webcam (no noise, blur or colour "
                  "effects, light JPEG).",
                  "- Ten opening moves from the start position: e7e5, d7d5, g8f6, b8c6, c7c5 (the robot's "
                  "side) and e2e4, d2d4, g1f3, b1c3, c2c4 (the far side).",
                  "- Markers with an opaque 2 px outline; pieces centred in their squares.",
                  "- Sharper video: H.264 at CRF 18 with full 4:4:4 chroma.", ""]
    if "baseline_recovery_v1" in names:
        lines += ["## Corrections (baseline_recovery_v1)", "",
                  "DAgger-style data in the baseline scene. The baseline policy "
                  "(`Machanize/chess_phase_smolvla_baseline`) drives; the scripted teacher takes over from "
                  "exactly where the arm is, and only the teacher's part is recorded. A grasp inspection had "
                  "shown the policy's jaws close 4-6 mm off the piece when it fails, against about 1 mm when "
                  "it holds. Handovers (`recovery.trigger` in `phase2_episodes.jsonl`):", "",
                  "- `pre_grasp`: the jaws, still open, come within 15-45 mm above the piece. The teacher "
                  "lines up above it, holds still 0.5 s (the wrist camera sees the piece between the jaws), "
                  "descends, grasps and finishes the move.",
                  "- `misaligned_close`: the policy starts closing more than 2.5 mm off. The teacher reopens, "
                  "rises clear, holds still 0.5 s to look, lines up, grasps and finishes.",
                  "- `failed_lift`: the jaws closed but the piece did not rise within 1.2 s. Same recovery.", "",
                  "Episodes where the piece had toppled or a neighbour had moved before the handover are "
                  "dropped, as are failed recoveries.", "",
                  "Fine-tuning the baseline on these plus baseline_v1 gave 20% (from 36.7%): in `pre_grasp` "
                  "and after a misaligned close the teacher first rises to its approach height and pauses, "
                  "and the policy learned to hover over pawns. The next attempt (`--style v2`) lines up from "
                  "where the arm is and goes straight down.", ""]
    if "baseline_dart_v1" in names:
        lines += ["## Perturbed teacher (baseline_dart_v1)", "",
                  "DART-style demonstrations (`PHASE2_PROFILE=baseline,dart`, `Expert.perturb_pick`). The plain "
                  "teacher always starts its descent exactly centred and goes straight down, so near the piece "
                  "its next action can be predicted from the joint state alone; the baseline policy's roughly "
                  "7 mm aiming error is never corrected on the way down.", "",
                  "- In about half the episodes the executed hand drifts sideways by a random 2-10 mm (any "
                  "direction) during the approach, holds the offset into the descent and returns to the plan "
                  "over 0.3-0.8 s, ending 4 mm above the target's top. Each offset is collision-checked with "
                  "the target included.",
                  "- The recorded `action` is the clean plan throughout; `observation.state` and the images "
                  "show the perturbed hand. The episodes therefore show \"from off-centre, command the centred "
                  "path\", and never an off-centre aim.",
                  "- Same 300 episode seeds as baseline_v1 (`--seeds-from`), so each episode repeats the same "
                  "scene, move and timing. `dart_offset_mm` in `phase2_episodes.jsonl` is the executed offset "
                  "(null when unperturbed).", "",
                  "Teacher validation, 100 scenes: 100% success perturbed and plain, no touches on the target "
                  "or its neighbours before the grasp (`reports/dart_teacher/`).", ""]
    if val:
        lines += [f"Expert validation (random layouts): {val['success_percent']}% of {val['trials']} trials "
                  f"(target above {val['target_percent']}%).", ""]
    lines += [
        "## Load", "",
        "LeRobot expects a dataset at the top of a folder, so download it first (replace `<name>`):", "",
        "```bash",
        f"hf download {repo_id} --repo-type dataset --include \"{folder}/datasets/<name>/*\" --local-dir playful",
        "```", "", "```python",
        "from lerobot.datasets.lerobot_dataset import LeRobotDataset",
        f"ds = LeRobotDataset(\"local/so101_chess_sim\", root=\"playful/{folder}/datasets/<name>\")",
        "```", "",
    ]
    return "\n".join(lines)


def push(repo_id: str, dataset: Path, name: str, folder: str = "chess-sim", delete_local: bool = False,
         dry_run: bool = False) -> str:
    from huggingface_hub import HfApi

    with tempfile.TemporaryDirectory(dir=dataset.parent) as tmp:
        stage = Path(tmp) / folder
        _tree(dataset, stage / "datasets" / name, skip={"images", ".cache"})
        videos = list(dataset.glob("review_*.mp4"))
        for v in videos:
            (stage / "datasets" / name / v.name).unlink()
            _place(v, stage / "review" / f"{name}.mp4")
        _tree(HERE / "reports", stage / "reports")
        _tree(HERE, stage / "code" / "sim", skip=SKIP_DIRS)
        _place(ROOT / "pyproject.toml", stage / "code" / "pyproject.toml")
        (stage / "README.md").write_text(readme(repo_id, folder, name, dataset))
        if dry_run:
            files = [p for p in Path(tmp).rglob("*") if p.is_file()]
            print((stage / "README.md").read_text())
            print(f"{len(files)} files, {sum(p.stat().st_size for p in files) / 1e6:.1f} MB, for example:")
            for p in sorted(files)[:: max(1, len(files) // 25)]:
                print("  ", p.relative_to(tmp))
            return "(dry run, nothing uploaded)"
        HfApi().upload_large_folder(repo_id=repo_id, folder_path=tmp, repo_type="dataset")
    if delete_local:
        shutil.rmtree(dataset)
    return f"https://huggingface.co/datasets/{repo_id}/tree/main/{folder}"


def main():
    from episode import load_config

    cfg = load_config()
    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("--repo-id", required=True, help="e.g. Machanize/playful")
    ap.add_argument("--name", required=True, help="dataset folder name, e.g. pilot_60")
    ap.add_argument("--root", default=cfg["dataset"]["root"])
    ap.add_argument("--path-in-repo", default="chess-sim")
    ap.add_argument("--delete-local", action="store_true", help="remove the local dataset after a successful upload")
    ap.add_argument("--dry-run", action="store_true", help="stage and list the files, upload nothing")
    args = ap.parse_args()
    url = push(args.repo_id, (ROOT / args.root).resolve(), args.name, args.path_in_repo, args.delete_local,
               args.dry_run)
    print(f"uploaded to {url}")


if __name__ == "__main__":
    main()