chess-sim/code/sim/jaw_tolerance.py

"""How much grasp error the jaws tolerate, at the current opening and at wider ones.

The scripted expert grasps pieces in the settings profile's scene (run with
PHASE2_PROFILE=baseline) on purpose off-centre: the pinch point is shifted 0, 2, 4 or
6 mm along the jaws' closing direction (both ways) or across it, with the jaws opened
3 mm beyond the piece on each side (the expert's current setting) or 5, 7, 9 mm. It
descends straight down from above, closes and lifts 30 mm. No collision checking: the
point is to see what physically happens.

Per trial: held (piece lifted 15 mm and upright), toppled (tilted over 45 deg), not
lifted, and whether another piece moved more than 2 mm (wider jaws can hit
neighbours) or the piece was pushed more than 2 mm before the jaws closed.

Writes jaw_tolerance.md and jaw_tolerance.json to --out.

Run:  PHASE2_PROFILE=baseline .venv/bin/python sim/jaw_tolerance.py
"""
from __future__ import annotations

import argparse
import json
import multiprocessing as mp
import sys
from collections import defaultdict
from pathlib import Path

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

SOURCES = {"e7": "e5", "g8": "f6", "e2": "e4", "b1": "c3"}     # pawn and knight, near and far side
MARGINS_MM = (3.0, 5.0, 7.0, 9.0)
OFFSETS_MM = (0.0, 2.0, 4.0, 6.0)
DIRECTIONS = ("along+", "along-", "across")


def trial(job):
    import warnings

    warnings.filterwarnings("ignore")
    import chess
    import mujoco
    import numpy as np

    from episode import EpisodeRunner, Task, load_config
    from expert import Destination, Trajectory
    from piece_sets import sample_piece_set

    global _runner
    if "_runner" not in globals():
        cfg = load_config()
        _runner = EpisodeRunner(cfg, sample_piece_set(np.random.default_rng(0), cfg, "jaw"), render=False)
    r = _runner
    ex, d, w = r.expert, r.d, r.w
    src, margin, off, direction, seed = job
    rng = np.random.default_rng(seed)
    squares = r._assign_bodies(chess.Board(), rng)
    task = Task(squares[src], src, Destination("square", None, SOURCES[src]), squares, [], chess.Board().fen(), "jaw")
    task = r.setup(np.random.default_rng(seed), task=task)
    name = task.target
    plan = ex.plan_pick(d, name, np.random.default_rng(seed))          # the expert's yaw for this piece
    grasp, offset = ex.grasp_for(name, plan.yaw, d, margin / 1000)
    lo, hi = r.grip_range
    if not lo <= grasp.open_q <= hi:
        return dict(job=job, outcome="opening beyond the gripper")
    axis = np.array([np.cos(plan.yaw), np.sin(plan.yaw), 0.0])
    across = np.array([-axis[1], axis[0], 0.0])
    vec = {"along+": axis, "along-": -axis, "across": across}[direction]
    base = w.base_pos(d, name)
    point = base + [0, 0, grasp.tip_height] + vec * off / 1000
    top = ex._neighbour_tops(d, base, 0.045, exclude=(name,))
    above = np.r_[point[:2], max(top + ex.e["travel_clearance_mm"] / 1000 + 0.004, point[2] + 0.025)]
    ex.speed = 1.0
    ex._grip_floor = grasp.contact_q
    traj = Trajectory()
    q_now = d.qpos[ex.kin.qadr].copy()
    q_above = ex._solve(above, plan.yaw, offset, plan.q_grasp, global_search=True)
    ex._joint_move(traj, q_now, q_above, float(d.qpos[r.grip_qadr]), grasp.open_q, "approach")
    q = ex._cart_move(traj, q_above, above, point, plan.yaw, plan.yaw, offset, grasp.open_q, grasp.open_q, "descend",
                      ex.e["vertical_speed"], 0.7, 1.8)
    n_before_close = len(traj)
    ex._hold(traj, q, grasp.open_q, ex.close_q, 0.45, "close")
    ex._hold(traj, q, ex.close_q, ex.close_q, 0.25, "close")
    ex._cart_move(traj, q, point, point + [0, 0, 0.03], plan.yaw, plan.yaw, offset, ex.close_q, ex.close_q, "lift",
                  ex.e["vertical_speed"], 0.5, 1.5)
    start = {n: w.base_pos(d, n) for n in w.pieces}
    others = [n for n in task.squares.values() if n != name]
    pushed = 0.0
    for i, (qq, g, ph) in enumerate(zip(traj.q, traj.g, traj.phase)):
        r._step_frame(qq, g, task, None, ph, lambda phase: None)
        if i == n_before_close - 1:
            pushed = float(np.linalg.norm(w.base_pos(d, name)[:2] - start[name][:2]))
    rise = float(w.base_pos(d, name)[2] - start[name][2])
    tilt = float(w.tilt_deg(d, name))
    moved = max(float(np.linalg.norm(w.base_pos(d, n) - start[n])) for n in others)
    outcome = ("toppled" if tilt > 45 else "held" if rise > 0.015 and tilt < 20 else "not lifted")
    return dict(job=job, piece=w.kind[name], outcome=outcome, neighbour_moved=moved > 0.002,
                pushed_before_close=pushed > 0.002, rise_mm=round(1000 * rise, 1), tilt_deg=round(tilt, 1),
                open_q_rad=round(float(grasp.open_q), 4))          # gripper joint angle, not a width


def main():
    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("--repeats", type=int, default=2, help="trials per condition (different start poses)")
    ap.add_argument("--workers", type=int, default=8)
    ap.add_argument("--out", default=str(HERE / "reports" / "baseline_eval" / "jaw_tolerance"))
    args = ap.parse_args()
    jobs = []
    for src in SOURCES:
        for margin in MARGINS_MM:
            for off in OFFSETS_MM:
                for direction in (("along+",) if off == 0 else DIRECTIONS):
                    for rep in range(args.repeats):
                        jobs.append((src, margin, off, direction, 1000 * rep + len(jobs)))
    with mp.get_context("spawn").Pool(args.workers) as pool:
        results = pool.map(trial, jobs, chunksize=4)
    out = Path(args.out)
    out.mkdir(parents=True, exist_ok=True)
    (out / "jaw_tolerance.json").write_text(json.dumps(results, indent=1))
    table = defaultdict(list)
    for r in results:
        src, margin, off, direction, _ = r["job"]
        piece = r.get("piece", "pawn" if src in ("e7", "e2") else "knight")
        table[(piece, margin, off)].append(r)
    lines = ["# Jaw tolerance", "",
             "The scripted expert grasps pieces in the baseline scene deliberately off-centre (pinch point shifted "
             "along the closing direction, both ways, or across it) with the jaws opened 3 mm beyond the piece on "
             "each side (current) or wider, then lifts 30 mm. **Clean hold** = lifted and upright with no other "
             "piece moved more than 2 mm.", "",
             f"{len(results)} trials, pawns on e7/e2 and knights on g8/b1, {args.repeats} start poses per condition.", ""]
    for piece in ("pawn", "knight"):
        lines += [f"## {piece.capitalize()}s: clean hold % (trials)", "",
                  "| opening per side | " + " | ".join(f"{o:.0f} mm off" for o in OFFSETS_MM) + " |",
                  "|---|" + "---|" * len(OFFSETS_MM)]
        for margin in MARGINS_MM:
            cells = []
            for off in OFFSETS_MM:
                rs = table[(piece, margin, off)]
                clean = sum(r["outcome"] == "held" and not r["neighbour_moved"] for r in rs)
                cells.append(f"{100 * clean / max(len(rs), 1):.0f} ({len(rs)})")
            lines.append(f"| {margin:.0f} mm{' (current)' if margin == 3 else ''} | " + " | ".join(cells) + " |")
        rs = [r for k, v in table.items() if k[0] == piece for r in v]
        lines += ["", "Failures by kind: " + ", ".join(
            f"{k} {v}" for k, v in sorted(defaultdict(int, {o: sum(r['outcome'] == o for r in rs)
                                                           for o in ('toppled', 'not lifted',
                                                                     'opening beyond the gripper')}).items())) +
                  f"; a neighbour moved in {sum(r.get('neighbour_moved', False) for r in rs)}; the piece was pushed "
                  f"before closing in {sum(r.get('pushed_before_close', False) for r in rs)}.", ""]
    (out / "jaw_tolerance.md").write_text("\n".join(lines) + "\n")
    print("\n".join(lines))


if __name__ == "__main__":
    main()