| """Поиск соглашения о повороте, в котором размечен BridgeData. |
| |
| Цель известна: norm_stats чекпойнта задают распределение roll/pitch/yaw, на |
| котором модель обучалась. Перебираем звенья робота, системы отсчёта и порядки |
| Эйлера и смотрим, какая комбинация укладывается в [q01, q99]. |
| |
| Модель не нужна — только среда и статистики. |
| """ |
| import json, os, sys, pathlib, itertools |
| import numpy as np |
|
|
| os.environ.setdefault("MUJOCO_GL", "egl") |
| import simpler_env |
| from scipy.spatial.transform import Rotation |
|
|
| sys.path.insert(0, "/root/greenvla_repo") |
| from lerobot.common.policies.factory import _resolve_norm_stats_dir |
|
|
| CKPT = "SberRoboticsCenter/GreenVLA-5b-stride-1-R1-bridge" |
| d = pathlib.Path(_resolve_norm_stats_dir(CKPT, "bridge")) / "bridge" / "norm_stats.json" |
| ns = json.load(open(d)) |
| ns = ns.get("norm_stats", ns) |
| q01 = np.array(ns["state"]["q01"], float)[:7] |
| q99 = np.array(ns["state"]["q99"], float)[:7] |
| mean = np.array(ns["state"]["mean"], float)[:7] |
| print("цель roll/pitch/yaw: q01=%s q99=%s" % ( |
| " ".join("%.3f" % x for x in q01[3:6]), " ".join("%.3f" % x for x in q99[3:6]))) |
|
|
| |
| env = simpler_env.make("widowx_spoon_on_towel") |
| poses = [] |
| rng = np.random.default_rng(0) |
| for ep in range(6): |
| obs, _ = env.reset(options={"obj_init_options": {"episode_id": ep}}) |
| agent = env.unwrapped.agent |
| for step in range(25): |
| root = agent.robot.get_root_pose() |
| snap = {} |
| for l in agent.robot.get_links(): |
| pl = l.get_pose() |
| snap[l.get_name()] = (np.array(pl.p, float), np.array(pl.q, float)) |
| poses.append((snap, (np.array(root.p, float), np.array(root.q, float)))) |
| a = np.zeros(7) |
| a[:3] = rng.normal(0, 0.03, 3) |
| a[3:6] = rng.normal(0, 0.10, 3) |
| a[6] = 1.0 if step % 10 < 5 else -1.0 |
| obs, r, term, trunc, info = env.step(a) |
| if term or trunc: |
| break |
|
|
| print("поз собрано: %d" % len(poses)) |
| names = sorted(poses[0][0].keys()) |
| print("звенья: %s" % names) |
|
|
| CAND_LINKS = [n for n in names if any(k in n.lower() for k in ("ee", "grip", "tcp", "wrist", "hand"))] |
| if not CAND_LINKS: |
| CAND_LINKS = names |
| ORDERS = ["xyz", "XYZ", "zyx", "ZYX", "xyx", "zxz"] |
|
|
|
|
| def score(rpy): |
| """Доля каналов, попавших в обучающий диапазон, + штраф за выход.""" |
| out = 0.0 |
| for i in range(3): |
| lo, hi = rpy[:, i].min(), rpy[:, i].max() |
| a_, b_ = q01[3 + i], q99[3 + i] |
| rng_ = max(b_ - a_, 1e-6) |
| out += max((a_ - lo) / rng_, (hi - b_) / rng_, 0.0) |
| return out |
|
|
|
|
| rows = [] |
| for link_name, base, order in itertools.product(CAND_LINKS, ["base", "world"], ORDERS): |
| try: |
| rpy = [] |
| for snap, root in poses: |
| lp, lq = snap[link_name] |
| R_l = Rotation.from_quat([lq[1], lq[2], lq[3], lq[0]]) |
| if base == "base": |
| rp, rq = root |
| R_r = Rotation.from_quat([rq[1], rq[2], rq[3], rq[0]]) |
| R_l = R_r.inv() * R_l |
| rpy.append(R_l.as_euler(order)) |
| rpy = np.array(rpy) |
| rows.append((score(rpy), link_name, base, order, rpy)) |
| except Exception: |
| continue |
|
|
| rows.sort(key=lambda r: r[0]) |
| print("") |
| print("%-26s%-7s%-6s%9s %s" % ("звено", "отсчёт", "поряд", "штраф", "roll/pitch/yaw min..max")) |
| for s, ln, base, order, rpy in rows[:14]: |
| desc = " ".join("[%.2f,%.2f]" % (rpy[:, i].min(), rpy[:, i].max()) for i in range(3)) |
| print("%-26s%-7s%-6s%9.2f %s" % (ln[:25], base, order, s, desc)) |
|
|
