evaluation_all / code /smoke_env.py
yqi19's picture
Upload folder using huggingface_hub
8aa2acf verified
Raw
History Blame Contribute Delete
1.82 kB
"""Env-only smoke test (no GR00T server): build VerbObjectColor-v1 via
groot_main._build_env_and_instruction for a few experiments, reset+step,
print obs/info structure. Validates the ManiSkill side + our kwargs + GPU sim.
"""
import sys, numpy as np
sys.path.insert(0, "/workspace/groot_eval/harness")
from groot_main import _build_env_and_instruction, _state8, _to_hwc_uint8, Args, SPATIAL_ANCHORS, SPATIALS
CASES = [
dict(experiment="color_object", pair_i=0, pair_j=1, run_type="color"),
dict(experiment="spatial_object", pair_i=0, pair_j=1, run_type="spatial"),
]
for c in CASES:
a = Args(sim_backend=sys.argv[1] if len(sys.argv) > 1 else "gpu",
max_episode_steps=40, third_seed=42, seed=42, **c)
print(f"\n==== {c} ====")
env, instr = _build_env_and_instruction(a)
print("instruction:", repr(instr))
if a.experiment in {"verb_spatial","color_spatial","spatial_size","spatial_object"}:
ai = list(SPATIAL_ANCHORS[SPATIALS[a.pair_i]]); aj = list(SPATIAL_ANCHORS[SPATIALS[a.pair_j]])
ropts = {"num_distractors": 1, "obj_xy": ai, "distractor_xy": [aj]}
else:
ropts = {"num_distractors": 1}
obs, _ = env.reset(seed=42, options=ropts)
sd = obs["sensor_data"]
print("sensor_data keys:", list(sd.keys()))
b = _to_hwc_uint8(sd["base_camera"]["rgb"]); h = _to_hwc_uint8(sd["hand_camera"]["rgb"])
print("base img", b.shape, b.dtype, "wrist img", h.shape, h.dtype)
s = _state8(env); print("state8", s.shape, s.dtype, np.round(s, 3))
act = np.zeros(8, np.float32)
for t in range(5):
obs, r, term, trunc, info = env.step(act)
print("info keys:", sorted(info.keys()))
for k in ("success", "success_first_axis", "success_second_axis"):
print(f" {k} = {info.get(k)}")
env.close()
print("\nSMOKE_ENV_OK")