File size: 10,206 Bytes
8aa2acf | 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 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | #!/usr/bin/env python3
"""
GR00T N1.7 full-factor eval — aligned 1:1 with the pi0.5 protocol in
eval_pi0_5/examples/maniskill_full_factor/main.py.
The ENVIRONMENT, cell vocabulary, instruction format, distractor / size /
spatial logic, success criterion and the `Success rate: X / Y (Z%)` stdout
line are COPIED VERBATIM from the pi0.5 harness so the numbers are directly
comparable. The only thing that differs is the policy boundary: instead of
the openpi websocket client we drive a fine-tuned GR00T N1.7 checkpoint
served over zmq by gr00t.eval.run_gr00t_server (same wire format the conflict
harness uses — see groot_main.py::_query_groot).
"""
from __future__ import annotations
import collections
import dataclasses
import logging
import pathlib
import random
import gymnasium as gym
import imageio.v2 as imageio
import mani_skill.envs # noqa: F401
import numpy as np
import torch
import tqdm
import tyro
from groot_client import GrootClient
# ── Vocabularies (VERBATIM from pi0.5 main.py) ────────────────────────────────
TRAINING_VERBS = ("lift", "grasp", "push", "pull", "rotate", "slide")
TRAINING_COLORS = ("red", "yellow", "blue", "orange", "green", "black")
TRAINING_SHAPES = ("cube", "sphere", "cup", "car", "pyramid", "star")
TRAINING_SPATIALS = ("left", "right", "middle", "front", "behind")
TRAINING_SIZES = ("small", "large", "smaller", "larger")
COLOR_TO_ID = {c: i for i, c in enumerate(TRAINING_COLORS)}
VERB_TO_EN = {
"lift": "Lift", "grasp": "Grasp", "push": "Push",
"pull": "Pull", "rotate": "Rotate", "slide": "Slide",
}
SPATIAL_TO_PHRASE = {
"left": "on the left", "right": "on the right", "middle": "in the middle",
"front": "in front", "behind": "at the back",
}
SPATIAL_XY_ANCHOR = {
"left": (-0.10, 0.00),
"right": ( 0.10, 0.00),
"middle": ( 0.00, 0.00),
"front": ( 0.00, 0.10),
"behind": ( 0.00, -0.10),
}
SIZE_CONFIG = {
"small": dict(target_size_scale=0.72, distractor_size_scales=None),
"large": dict(target_size_scale=1.34, distractor_size_scales=None),
"smaller": dict(target_size_scale=0.82, distractor_size_scales=[1.08]),
"larger": dict(target_size_scale=1.18, distractor_size_scales=[0.92]),
}
def make_instruction(verb: str, size: str, color: str, shape: str, spatial: str) -> str:
return f"{VERB_TO_EN[verb]} the {size} {color} {shape} {SPATIAL_TO_PHRASE[spatial]}."
@dataclasses.dataclass
class Args:
# GR00T zmq policy server
host: str = "127.0.0.1"
port: int = 5555
replan_steps: int = 5
# Task specification (5 factors)
verb: str = "lift"
color: str = "red"
shape: str = "cube"
spatial: str = "left"
size: str = "small"
prompt: str = ""
"""Override language instruction; if empty, auto-built from the 5 factors."""
no_distractor_prob: float = 0.70
"""Probability per episode of forcing num_distractors=0 (pi0.5: 0.70)."""
# ManiSkill
num_episodes: int = 50
max_episode_steps: int = 500
sim_backend: str = "cpu"
render_backend: str = "cpu"
obs_mode: str = "rgb"
render_mode: str | None = None
seed: int = 0
# Output
video_out_path: str = "data/maniskill_full_factor/videos"
save_wrist_video: bool = True
# ── Helpers (VERBATIM from pi0.5 main.py) ─────────────────────────────────────
def _to_numpy_hwc(x: np.ndarray | torch.Tensor) -> np.ndarray:
if torch.is_tensor(x):
x = x.detach().float().cpu().numpy()
x = np.asarray(x)
if x.ndim == 4:
x = x[0]
if x.shape[0] in (1, 3) and x.shape[-1] != 3 and x.ndim == 3:
x = np.transpose(x, (1, 2, 0))
if np.issubdtype(x.dtype, np.floating) and x.max() <= 1.0:
x = (np.clip(x, 0, 1) * 255).astype(np.uint8)
else:
x = x.astype(np.uint8)
return np.ascontiguousarray(x)
def _state8(env: gym.Env) -> np.ndarray:
qpos = env.unwrapped.agent.robot.get_qpos()
if torch.is_tensor(qpos):
qpos = qpos[0].detach().cpu().numpy()
qpos = np.asarray(qpos, dtype=np.float32).ravel()
if qpos.size >= 8:
return qpos[:8].copy()
out = np.zeros(8, dtype=np.float32)
out[: qpos.size] = qpos
return out
def _success(info: dict) -> bool:
if "success" not in info:
return False
s = info["success"]
if torch.is_tensor(s):
return bool(s.squeeze().item())
return bool(np.asarray(s).squeeze())
def _spatial_xy(spatial: str, rng: random.Random) -> list[float]:
ax, ay = SPATIAL_XY_ANCHOR[spatial]
return [ax + rng.uniform(-0.012, 0.012), ay + rng.uniform(-0.012, 0.012)]
# ── GR00T policy boundary (VERBATIM from groot_main.py::_query_groot) ──────────
def _query_groot(client: GrootClient, img_base: np.ndarray, img_wrist: np.ndarray,
state8: np.ndarray, instruction: str) -> np.ndarray:
"""Build the nested GR00T observation, query the server, return an
(action_horizon, 8) float32 chunk = [7 joint-pos targets, 1 gripper]."""
obs = {
"video": {
"image": img_base[None, None, ...],
"wrist_image": img_wrist[None, None, ...],
},
"state": {
"arm": state8[:7][None, None, :].astype(np.float32),
"gripper": state8[7:8][None, None, :].astype(np.float32),
},
"language": {
"annotation.human.task_description": [[instruction]],
},
}
action, _info = client.get_action(obs)
arm = np.asarray(action["arm"], dtype=np.float32) # (1, Th, 7)
grip = np.asarray(action["gripper"], dtype=np.float32) # (1, Th, 1)
return np.concatenate([arm[0], grip[0]], axis=-1) # (Th, 8)
# ── Main eval loop (env / sampling logic VERBATIM from pi0.5 main.py) ──────────
def eval_full_factor(args: Args) -> None:
logging.basicConfig(level=logging.INFO, force=True)
verb = args.verb.lower().strip()
color = args.color.lower().strip()
shape = args.shape.lower().strip()
spatial = args.spatial.lower().strip()
size = args.size.lower().strip()
for val, vocab, name in [
(verb, TRAINING_VERBS, "verb"),
(color, TRAINING_COLORS, "color"),
(shape, TRAINING_SHAPES, "shape"),
(spatial, TRAINING_SPATIALS, "spatial"),
(size, TRAINING_SIZES, "size"),
]:
if val not in vocab:
raise ValueError(f"{name}={val!r} not in {vocab}")
prompt = args.prompt.strip() or make_instruction(verb, size, color, shape, spatial)
logging.info("prompt=%r", prompt)
size_cfg = SIZE_CONFIG[size]
object_color_id = COLOR_TO_ID[color]
has_comparison = size_cfg["distractor_size_scales"] is not None
distractor_max = 1 if has_comparison else 0
make_kw: dict = dict(
obs_mode=args.obs_mode,
control_mode="pd_joint_pos",
sim_backend=args.sim_backend,
render_backend=args.render_backend,
max_episode_steps=args.max_episode_steps,
verb=verb,
object_shape=shape,
object_color_id=object_color_id,
distractor_max=distractor_max,
object_size_jiggle=0.0,
)
if args.render_mode is not None:
make_kw["render_mode"] = args.render_mode
env = gym.make("VerbObjectColor-v1", **make_kw)
video_out_path = pathlib.Path(args.video_out_path)
video_out_path.mkdir(parents=True, exist_ok=True)
client = GrootClient(args.host, args.port)
rng = random.Random(args.seed)
successes = 0
for ep in tqdm.tqdm(range(args.num_episodes)):
no_distractor = rng.random() < args.no_distractor_prob
reset_options: dict = {
"obj_xy": _spatial_xy(spatial, rng),
"target_size_scale": size_cfg["target_size_scale"],
}
if size_cfg["distractor_size_scales"] is not None:
reset_options["distractor_size_scales"] = size_cfg["distractor_size_scales"]
if no_distractor:
reset_options["num_distractors"] = 0
obs, _ = env.reset(seed=args.seed + ep, options=reset_options)
client.reset()
plan: collections.deque = collections.deque()
base_path = video_out_path / f"ep{ep:03d}.mp4"
wrist_path = video_out_path / f"ep{ep:03d}_wrist.mp4"
writer = imageio.get_writer(base_path, fps=30)
wrist_writer = imageio.get_writer(wrist_path, fps=30) if args.save_wrist_video else None
done = False
ep_success = False
try:
while not done:
rgb_b = _to_numpy_hwc(obs["sensor_data"]["base_camera"]["rgb"])
rgb_h = _to_numpy_hwc(obs["sensor_data"]["hand_camera"]["rgb"])
writer.append_data(rgb_b)
if wrist_writer is not None:
wrist_writer.append_data(rgb_h)
if not plan:
st = _state8(env)
chunk = _query_groot(client, rgb_b, rgb_h, st, prompt)
n = min(args.replan_steps, len(chunk))
if n < 1:
logging.warning("Empty action chunk from policy")
break
plan.extend(chunk[:n])
action = np.asarray(plan.popleft(), dtype=np.float32).ravel()[:8]
obs, _reward, term, trunc, info = env.step(action)
if _success(info):
ep_success = True
done = bool(term or trunc) or ep_success
finally:
try:
writer.close()
finally:
if wrist_writer is not None:
wrist_writer.close()
if ep_success:
successes += 1
logging.info("Episode %d success=%s no_distractor=%s", ep, ep_success, no_distractor)
env.close()
rate = successes / max(args.num_episodes, 1)
logging.info("Success rate: %d / %d (%.1f%%)", successes, args.num_episodes, 100.0 * rate)
def main() -> None:
eval_full_factor(tyro.cli(Args))
if __name__ == "__main__":
main()
|