VLAarchtests / tests /test_proxy_scripted_bench.py
lsnu's picture
2026-03-25 runpod handoff update
e7d8e79 verified
from sim_reveal.procedural_envs import BAG_PROXY, CLOTH_PROXY, FOLIAGE_PROXY, make_proxy_env
def test_proxy_scripted_bench():
bag_wins = 0
foliage_wins = 0
cloth_wins = 0
for seed in range(10):
bag_env = make_proxy_env(proxy_name=BAG_PROXY.name, resolution=32, seed=seed, rollout_horizon=4)
bag_env.reset(seed=seed)
maintain = bag_env.evaluate_action_chunk(bag_env.macro_action_chunk("maintain_mouth", chunk_horizon=4), rollout_horizon=4)
retrieve = bag_env.evaluate_action_chunk(bag_env.macro_action_chunk("premature_retrieve", chunk_horizon=4), rollout_horizon=4)
if (maintain["hold_persistence"] >= retrieve["hold_persistence"]) and (maintain["reocclusion_rate"] <= retrieve["reocclusion_rate"]):
bag_wins += 1
foliage_env = make_proxy_env(proxy_name=FOLIAGE_PROXY.name, resolution=32, seed=seed, rollout_horizon=4)
foliage_env.reset(seed=seed)
pin = foliage_env.evaluate_action_chunk(foliage_env.macro_action_chunk("pin_canopy", chunk_horizon=4), rollout_horizon=4)
swipe = foliage_env.evaluate_action_chunk(foliage_env.macro_action_chunk("foliage_immediate_reocclusion", chunk_horizon=4), rollout_horizon=4)
if (pin["reocclusion_rate"] <= swipe["reocclusion_rate"]) and (pin["visibility_integral"] >= swipe["visibility_integral"]):
foliage_wins += 1
cloth_env = make_proxy_env(proxy_name=CLOTH_PROXY.name, resolution=32, seed=seed, rollout_horizon=4)
cloth_env.reset(seed=seed)
stabilize = cloth_env.evaluate_action_chunk(cloth_env.macro_action_chunk("stabilize_fold", chunk_horizon=4), rollout_horizon=4)
lift_high = cloth_env.evaluate_action_chunk(cloth_env.macro_action_chunk("cloth_lift_high", chunk_horizon=4), rollout_horizon=4)
if (stabilize["candidate_fold_preservation"] >= lift_high["candidate_fold_preservation"]) and (
stabilize["final_disturbance_cost"] <= lift_high["final_disturbance_cost"]
):
cloth_wins += 1
assert bag_wins >= 8
assert foliage_wins >= 8
assert cloth_wins >= 8