evaluation_all / code /groot_grid_eval.py
yqi19's picture
Upload folder using huggingface_hub
8aa2acf verified
#!/usr/bin/env python3
"""
GR00T PAIRWISE-GRID eval (HARD口径) — mirrors pi0_grid_eval.py exactly, with
the policy boundary swapped from openpi-websocket to GR00T zmq (GrootClient).
Same HARD color_size / color_spatial cell construction; same env reset
options; same per-cell scoring & results-txt format; same 3-seed protocol.
"""
from __future__ import annotations
import argparse
import os
import pathlib
import random
import sys
import numpy as np
MGEN_ROOT = os.environ.get("MGEN_ROOT", "/workspace/Maniskill_gen_new")
SIM_ROOT = os.environ.get("SIM_ROOT", "/workspace/eval_simulation/simulation")
HARN_DIR = pathlib.Path(__file__).resolve().parent
for _p in (str(HARN_DIR), SIM_ROOT, MGEN_ROOT):
if _p not in sys.path:
sys.path.insert(0, _p)
import gymnasium as gym # noqa: E402
import mani_skill.envs # noqa: E402,F401 (registers VerbObjectColor-v1)
from groot_client import GrootClient # noqa: E402
# Repo's own canonical env-id/color resolver (handles legacy routing exactly).
from scripts.run_verb_color_shape_motion_planning import get_env_id_and_color # noqa: E402
COLORS = ("red", "yellow", "blue", "orange", "green", "black")
SIZES = ("small", "large", "smaller", "larger", "smallest", "largest")
SPATIALS = ("left", "right", "middle", "front", "behind")
SHAPES = ("cube", "sphere", "cup", "car", "pyramid", "star")
VERB_POOL = ("lift", "grasp", "push")
VERB_CAP = {"lift": "Lift", "grasp": "Grasp", "push": "Push",
"pull": "Pull", "rotate": "Rotate", "slide": "Slide"}
SPATIAL_PHRASE = {"left": "on the left", "right": "on the right",
"middle": "in the middle", "front": "in front",
"behind": "at the back"}
SPATIAL_ANCHOR = {"left": (-0.10, 0.0), "right": (0.10, 0.0),
"middle": (0.0, 0.0), "front": (0.0, 0.10),
"behind": (0.0, -0.10)}
def _spatial_xy(spatial, rng):
ax, ay = SPATIAL_ANCHOR[spatial]
return [ax + rng.uniform(-0.012, 0.012), ay + rng.uniform(-0.012, 0.012)]
def _to_hwc_uint8(x):
import torch
if torch.is_tensor(x):
x = x.detach().float().cpu().numpy()
x = np.asarray(x)
if x.ndim == 4:
x = x[0]
if x.ndim == 3 and x.shape[0] in (1, 3) and x.shape[-1] != 3:
x = np.transpose(x, (1, 2, 0))
if np.issubdtype(x.dtype, np.floating) and x.max() <= 1.0 + 1e-6:
x = (np.clip(x, 0, 1) * 255).astype(np.uint8)
else:
x = x.astype(np.uint8)
return np.ascontiguousarray(x)
def _state8(env):
import torch
q = env.unwrapped.agent.robot.get_qpos()
if torch.is_tensor(q):
q = q[0].detach().cpu().numpy()
q = np.asarray(q, dtype=np.float32).ravel()
out = np.zeros(8, dtype=np.float32)
out[: min(8, q.size)] = q[: min(8, q.size)]
return out
def _success(info):
import torch
s = info.get("success", False)
if torch.is_tensor(s):
return bool(s.squeeze().item())
return bool(np.asarray(s).squeeze())
# HARD口径 — same as pi0_grid_eval.HARD_SIZE
HARD_SIZE = {
"small": (0.72, [1.20], 1),
"large": (1.34, [0.80], 1),
"smaller": (0.82, [1.08], 1),
"larger": (1.18, [0.92], 1),
"smallest": (0.78, [1.00, 1.24], 2),
"largest": (1.26, [1.00, 0.80], 2),
}
# EASY口径 — 1 distractor only, ~2x contrast, drops smallest/largest double-distractor
EASY_SIZE = {
"small": (0.65, [1.35], 1),
"large": (1.40, [0.65], 1),
"smaller": (0.70, [1.30], 1),
"larger": (1.30, [0.70], 1),
"smallest": (0.65, [1.40], 1),
"largest": (1.40, [0.65], 1),
}
def build_color_size_cell(color, size_label, *, distractor_max_arg, task_difficulty, mode="hard"):
verb, shape = "lift", "cube"
table = HARD_SIZE if mode == "hard" else EASY_SIZE
t_scale, d_scales, n_d = table[size_label]
env_id, color_id, extra = get_env_id_and_color(
verb, color, shape, distractor_max=max(int(distractor_max_arg), n_d),
task_difficulty=task_difficulty)
make_kw = dict(obs_mode="rgb", control_mode="pd_joint_pos",
render_mode="rgb_array")
if env_id == "VerbObjectColor-v1":
make_kw.update(extra)
make_kw["object_size_jiggle"] = 0.0
make_kw["distractor_specs"] = [("cube", int(color_id))] * n_d + \
[None] * (3 - n_d)
else:
make_kw["object_color_id"] = color_id
reset_opts = {"num_distractors": int(n_d),
"target_size_scale": float(t_scale),
"distractor_size_scales": [float(x) for x in d_scales]}
instruction = f"Lift the {size_label} {color} cube."
return env_id, make_kw, reset_opts, instruction
def build_color_spatial_cell(color, spatial, *, rng, distractor_max_arg, task_difficulty):
verb = rng.choice(VERB_POOL)
shape = "cube"
others = [s for s in SPATIALS if s != spatial]
d_spatial = rng.choice(others)
env_id, color_id, extra = get_env_id_and_color(
verb, color, shape, distractor_max=1, task_difficulty=task_difficulty)
make_kw = dict(obs_mode="rgb", control_mode="pd_joint_pos",
render_mode="rgb_array")
if env_id == "VerbObjectColor-v1":
make_kw.update(extra)
make_kw["object_size_jiggle"] = 0.0
make_kw["distractor_specs"] = [("cube", int(color_id)), None, None]
else:
make_kw["object_color_id"] = color_id
reset_opts = {"num_distractors": 1, "target_size_scale": 1.0,
"obj_xy": _spatial_xy(spatial, rng),
"distractor_xy": [_spatial_xy(d_spatial, rng)],
"distractor_size_scales": [1.0]}
instruction = f"{VERB_CAP[verb]} the {color} cube {SPATIAL_PHRASE[spatial]}."
return env_id, make_kw, reset_opts, instruction
def build_verb_spatial_cell(verb, spatial, *, rng, distractor_max_arg, task_difficulty):
"""HARD: 2 same-color same-shape(cube) cubes at DIFFERENT spatial anchors.
Color sampled per-cell (third factor); instruction uses verb + spatial."""
color = rng.choice(COLORS)
shape = "cube"
others = [s for s in SPATIALS if s != spatial]
d_spatial = rng.choice(others)
env_id, color_id, extra = get_env_id_and_color(
verb, color, shape, distractor_max=1, task_difficulty=task_difficulty)
make_kw = dict(obs_mode="rgb", control_mode="pd_joint_pos",
render_mode="rgb_array")
if env_id == "VerbObjectColor-v1":
make_kw.update(extra)
make_kw["object_size_jiggle"] = 0.0
make_kw["distractor_specs"] = [("cube", int(color_id)), None, None]
else:
make_kw["object_color_id"] = color_id
reset_opts = {"num_distractors": 1, "target_size_scale": 1.0,
"obj_xy": _spatial_xy(spatial, rng),
"distractor_xy": [_spatial_xy(d_spatial, rng)],
"distractor_size_scales": [1.0]}
instruction = f"{VERB_CAP[verb]} the cube {SPATIAL_PHRASE[spatial]}."
return env_id, make_kw, reset_opts, instruction
def build_spatial_object_cell(spatial, shape, *, rng, distractor_max_arg, task_difficulty):
"""HARD: target shape at spatial; distractor = DIFFERENT shape, same color,
at a DIFFERENT spatial. Instruction uses shape + spatial phrase."""
color = rng.choice(COLORS)
verb = rng.choice(VERB_POOL)
d_shape = rng.choice([s for s in SHAPES if s != shape])
d_spatial = rng.choice([s for s in SPATIALS if s != spatial])
env_id, color_id, extra = get_env_id_and_color(
verb, color, shape, distractor_max=1, task_difficulty=task_difficulty)
make_kw = dict(obs_mode="rgb", control_mode="pd_joint_pos",
render_mode="rgb_array")
if env_id == "VerbObjectColor-v1":
make_kw.update(extra)
make_kw["object_size_jiggle"] = 0.0
make_kw["distractor_specs"] = [(d_shape, int(color_id)), None, None]
else:
make_kw["object_color_id"] = color_id
reset_opts = {"num_distractors": 1, "target_size_scale": 1.0,
"obj_xy": _spatial_xy(spatial, rng),
"distractor_xy": [_spatial_xy(d_spatial, rng)],
"distractor_size_scales": [1.0]}
instruction = f"{VERB_CAP[verb]} the {shape} {SPATIAL_PHRASE[spatial]}."
return env_id, make_kw, reset_opts, instruction
def build_verb_size_cell(verb, size_label, *, rng, distractor_max_arg, task_difficulty):
"""HARD: target cube of size; distractor(s) = same-color cube differing
ONLY in size. Instruction uses verb + size word."""
color = rng.choice(COLORS)
shape = "cube"
t_scale, d_scales, n_d = HARD_SIZE[size_label]
env_id, color_id, extra = get_env_id_and_color(
verb, color, shape, distractor_max=max(int(distractor_max_arg), n_d),
task_difficulty=task_difficulty)
make_kw = dict(obs_mode="rgb", control_mode="pd_joint_pos",
render_mode="rgb_array")
if env_id == "VerbObjectColor-v1":
make_kw.update(extra)
make_kw["object_size_jiggle"] = 0.0
make_kw["distractor_specs"] = [("cube", int(color_id))] * n_d + \
[None] * (3 - n_d)
else:
make_kw["object_color_id"] = color_id
reset_opts = {"num_distractors": int(n_d),
"target_size_scale": float(t_scale),
"distractor_size_scales": [float(x) for x in d_scales]}
instruction = f"{VERB_CAP[verb]} the {size_label} cube."
return env_id, make_kw, reset_opts, instruction
def build_spatial_size_cell(spatial, size_label, *, rng, distractor_max_arg, task_difficulty):
"""HARD: 2 same-color same-shape(cube) cubes at DIFFERENT spatial AND
DIFFERENT size. Instruction uses both spatial phrase + size word."""
color = rng.choice(COLORS)
verb = rng.choice(VERB_POOL)
shape = "cube"
other_spatials = [s for s in SPATIALS if s != spatial]
d_spatial = rng.choice(other_spatials)
t_scale, _, _ = HARD_SIZE[size_label]
# distractor: opposite-extreme scale
if size_label in ("small", "smaller", "smallest"):
d_scale = 1.25
else:
d_scale = 0.75
env_id, color_id, extra = get_env_id_and_color(
verb, color, shape, distractor_max=1, task_difficulty=task_difficulty)
make_kw = dict(obs_mode="rgb", control_mode="pd_joint_pos",
render_mode="rgb_array")
if env_id == "VerbObjectColor-v1":
make_kw.update(extra)
make_kw["object_size_jiggle"] = 0.0
make_kw["distractor_specs"] = [("cube", int(color_id)), None, None]
else:
make_kw["object_color_id"] = color_id
reset_opts = {"num_distractors": 1,
"target_size_scale": float(t_scale),
"obj_xy": _spatial_xy(spatial, rng),
"distractor_xy": [_spatial_xy(d_spatial, rng)],
"distractor_size_scales": [float(d_scale)]}
instruction = f"{VERB_CAP[verb]} the {size_label} cube {SPATIAL_PHRASE[spatial]}."
return env_id, make_kw, reset_opts, instruction
def _groot_action_chunk(client: GrootClient, img_base, img_wrist, state8, instruction):
"""Build GR00T obs, call get_action, return (Th, 8) float32 chunk."""
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) or (Th, 7)
grip = np.asarray(action["gripper"], dtype=np.float32) # (1, Th, 1) or (Th, 1)
if arm.ndim == 3:
arm = arm[0]
if grip.ndim == 3:
grip = grip[0]
if grip.ndim == 1:
grip = grip[:, None]
return np.concatenate([arm, grip], axis=-1).astype(np.float32) # (Th, 8)
def run_cell(client, env_id, make_kw, reset_opts, instruction, *,
seed, sim_backend, max_steps, replan):
mk = dict(make_kw)
mk["sim_backend"] = sim_backend
mk["render_backend"] = sim_backend
env = gym.make(env_id, **mk)
obs, _ = env.reset(seed=seed, options=reset_opts)
plan = []
done = ok = False
steps = 0
try:
while not done and steps < max_steps:
b = _to_hwc_uint8(obs["sensor_data"]["base_camera"]["rgb"])
h = _to_hwc_uint8(obs["sensor_data"]["hand_camera"]["rgb"])
if not plan:
chunk = _groot_action_chunk(client, b, h, _state8(env), instruction)
n = min(replan, len(chunk))
if n < 1:
break
plan = list(chunk[:n])
act = np.asarray(plan.pop(0), dtype=np.float32).ravel()[:8]
obs, _r, term, trunc, info = env.step(act)
steps += 1
if _success(info):
ok = True
done = bool(term or trunc) or ok
finally:
env.close()
return ok
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--experiment", required=True,
choices=["color_size", "color_spatial", "verb_spatial",
"spatial_size", "spatial_object", "verb_size"])
ap.add_argument("--host", default="127.0.0.1")
ap.add_argument("--port", type=int, default=5600)
ap.add_argument("--seed", type=int, default=42)
ap.add_argument("--results-txt", required=True)
ap.add_argument("--sim-backend", default="gpu")
ap.add_argument("--task-difficulty", type=float, default=1.5)
ap.add_argument("--max-episode-steps", type=int, default=300)
ap.add_argument("--replan-steps", type=int, default=5)
ap.add_argument("--distractor-max", type=int, default=2)
ap.add_argument("--max-cells", type=int, default=0)
ap.add_argument("--target-episodes", type=int, default=0)
ap.add_argument("--color-size-mode", choices=["hard", "easy"], default="hard")
a = ap.parse_args()
if a.experiment == "color_size":
f1vals, f1name = COLORS, "color"
f2vals, f2name = SIZES, "size"
meta = f"fixed verb=lift shape=cube {a.color_size_mode.upper()} same-color cube distractor"
elif a.experiment == "color_spatial":
f1vals, f1name = COLORS, "color"
f2vals, f2name = SPATIALS, "spatial"
meta = "fixed shape=cube; verb~{lift,grasp,push}; HARD same-color cube distractor"
elif a.experiment == "verb_spatial":
f1vals, f1name = VERB_POOL, "verb"
f2vals, f2name = SPATIALS, "spatial"
meta = "fixed shape=cube; color~{COLORS}; HARD same-color cube distractor at different spatial"
elif a.experiment == "spatial_size":
f1vals, f1name = SPATIALS, "spatial"
f2vals, f2name = SIZES, "size"
meta = "fixed shape=cube; color&verb~random; HARD same-color cube distractor at different spatial & different size"
elif a.experiment == "spatial_object":
f1vals, f1name = SPATIALS, "spatial"
f2vals, f2name = SHAPES, "shape"
meta = "color&verb~random; HARD same-color different-shape distractor at different spatial"
else: # verb_size
f1vals, f1name = VERB_POOL, "verb"
f2vals, f2name = SIZES, "size"
meta = "fixed shape=cube; color~random; HARD same-color cube distractor differing only in size"
cells = [(c, x) for c in f1vals for x in f2vals]
if a.max_cells > 0:
cells = cells[: a.max_cells]
client = GrootClient(a.host, a.port)
# one ping to fail fast if server isn't up
if not client.ping():
raise SystemExit(f"GR00T server not reachable at {a.host}:{a.port}")
rt = pathlib.Path(a.results_txt)
rt.parent.mkdir(parents=True, exist_ok=True)
with rt.open("w") as f:
f.write(f"# gr00t {a.experiment} grid {meta} "
f"seed={a.seed} task_difficulty={a.task_difficulty} "
f"sim={a.sim_backend} cells={len(cells)}\n")
f.write(f"idx {f1name} {f2name} success prompt\n")
import math as _m
ncells = len(cells)
reps = max(1, _m.ceil(a.target_episodes / ncells)) if a.target_episodes > 0 else 1
print(f"cells={ncells} reps/cell={reps}{ncells*reps} episodes/seed", flush=True)
succ = tot = 0
for idx, (f1, f2) in enumerate(cells, 1):
for r in range(reps):
rng = random.Random(a.seed * 100003 + idx * 131 + r)
try:
if a.experiment == "color_size":
env_id, mk, ro, instr = build_color_size_cell(
f1, f2, distractor_max_arg=a.distractor_max,
task_difficulty=a.task_difficulty, mode=a.color_size_mode)
elif a.experiment == "color_spatial":
env_id, mk, ro, instr = build_color_spatial_cell(
f1, f2, rng=rng, distractor_max_arg=a.distractor_max,
task_difficulty=a.task_difficulty)
elif a.experiment == "verb_spatial":
env_id, mk, ro, instr = build_verb_spatial_cell(
f1, f2, rng=rng, distractor_max_arg=a.distractor_max,
task_difficulty=a.task_difficulty)
elif a.experiment == "spatial_size":
env_id, mk, ro, instr = build_spatial_size_cell(
f1, f2, rng=rng, distractor_max_arg=a.distractor_max,
task_difficulty=a.task_difficulty)
elif a.experiment == "spatial_object":
env_id, mk, ro, instr = build_spatial_object_cell(
f1, f2, rng=rng, distractor_max_arg=a.distractor_max,
task_difficulty=a.task_difficulty)
else: # verb_size
env_id, mk, ro, instr = build_verb_size_cell(
f1, f2, rng=rng, distractor_max_arg=a.distractor_max,
task_difficulty=a.task_difficulty)
ok = run_cell(client, env_id, mk, ro, instr,
seed=a.seed + idx * 1000 + r,
sim_backend=a.sim_backend,
max_steps=a.max_episode_steps, replan=a.replan_steps)
except Exception as e: # noqa: BLE001
print(f"[{idx}/{ncells} r{r}] FAIL {f1},{f2}: {e}", flush=True)
ok, instr = False, f"ERROR:{e}"
succ += int(ok)
tot += 1
with rt.open("a") as f:
f.write(f'{idx} {f1} {f2} rep{r} {int(ok)} "{instr}"\n')
print(f"[{idx}/{ncells}] {f1} {f2} {reps}reps → cum {succ}/{tot}",
flush=True)
rate = 100.0 * succ / tot if tot else 0.0
with rt.open("a") as f:
f.write(f"\noverall_success={succ}/{tot} ({rate:.1f}%)\n")
print(f"\nDone {a.experiment}: overall_success={succ}/{tot} ({rate:.1f}%)")
if __name__ == "__main__":
main()