Rocket_env / test_rocketenv.py
sarahzephyr's picture
Upload 10 files
f5a258c verified
Raw History Blame Contribute Delete
9.38 kB
"""
RocketEnv Tests β€” validates all OpenEnv spec requirements
"""
import pytest
from rocket_env import (
RocketEnv, RocketAction, RocketObservation,
ActionType, LaunchPhase, SubsystemStatus, TASK_CONFIGS
)
# ─────────────────────────────────────────────────────────────────────────────
# Reset tests
# ─────────────────────────────────────────────────────────────────────────────
def test_reset_returns_observation():
env = RocketEnv(task="nominal_launch", seed=42)
obs = env.reset()
assert isinstance(obs, RocketObservation)
def test_reset_produces_clean_state():
env = RocketEnv(task="nominal_launch", seed=42)
obs = env.reset()
assert obs.step_number == 0
assert obs.phase == LaunchPhase.PRE_CHECKS
assert obs.abort_risk == 0.0
assert len(obs.subsystems) == 6
def test_reset_deterministic_with_seed():
env = RocketEnv(task="nominal_launch", seed=123)
obs1 = env.reset()
env2 = RocketEnv(task="nominal_launch", seed=123)
obs2 = env2.reset()
assert obs1.subsystems == obs2.subsystems
assert obs1.weather_window == obs2.weather_window
def test_reset_all_tasks():
for task in TASK_CONFIGS:
env = RocketEnv(task=task, seed=0)
obs = env.reset()
assert obs is not None
assert obs.t_minus == TASK_CONFIGS[task]["t_minus_start"]
# ─────────────────────────────────────────────────────────────────────────────
# Step tests
# ─────────────────────────────────────────────────────────────────────────────
def test_step_returns_correct_types():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
result = env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
assert isinstance(result.reward, float)
assert isinstance(result.done, bool)
assert isinstance(result.observation, RocketObservation)
assert isinstance(result.info, dict)
def test_step_reward_in_range():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
for _ in range(10):
if env._done:
break
result = env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
assert -1.0 <= result.reward <= 1.0
def test_step_after_done_raises():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
# Force done state
env._done = True
with pytest.raises(RuntimeError):
env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
def test_diagnose_action():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
# Find a faulted subsystem
faulted = None
for name, sub in env._state["subsystems"].items():
if sub.status != SubsystemStatus.NOMINAL:
faulted = name
break
if faulted:
result = env.step(RocketAction(action_type=ActionType.DIAGNOSE, target=faulted))
assert result.reward > 0 # should reward fault discovery
def test_run_diagnostic_gives_positive_reward():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
result = env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
assert result.reward >= 0.0 # diagnostic always gives small positive reward
def test_override_penalized():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
result = env.step(RocketAction(action_type=ActionType.OVERRIDE, target="propulsion"))
assert result.reward < 0 # override should always penalize
def test_invalid_target_penalized():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
result = env.step(RocketAction(action_type=ActionType.DIAGNOSE, target="nonexistent"))
assert result.reward < 0
def test_step_increments_step_count():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
result = env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
assert result.observation.step_number == 1
# ─────────────────────────────────────────────────────────────────────────────
# State tests
# ─────────────────────────────────────────────────────────────────────────────
def test_state_returns_dict():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
s = env.state()
assert isinstance(s, dict)
assert "phase" in s
assert "subsystems" in s
assert "t_minus" in s
def test_state_before_reset():
env = RocketEnv(task="nominal_launch", seed=42)
s = env.state()
assert s["status"] == "not_initialized"
# ─────────────────────────────────────────────────────────────────────────────
# Grader tests
# ─────────────────────────────────────────────────────────────────────────────
def test_grader_returns_float_in_range():
for task in TASK_CONFIGS:
env = RocketEnv(task=task, seed=42)
env.reset()
# Run a few steps
for _ in range(5):
if env._done:
break
env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
score = env.grade()
assert isinstance(score, float), f"grade() must return float for task {task}"
assert 0.0 <= score <= 1.0, f"score out of range for task {task}: {score}"
def test_grader_higher_for_authorized_launch():
env = RocketEnv(task="nominal_launch", seed=42)
env.reset()
# Manually repair all faults
for name, sub in env._state["subsystems"].items():
sub.status = SubsystemStatus.NOMINAL
sub.fault_code = None
# Authorize
env.step(RocketAction(action_type=ActionType.AUTHORIZE))
score_good = env.grade()
env2 = RocketEnv(task="nominal_launch", seed=42)
env2.reset()
# Just abort immediately
env2.step(RocketAction(action_type=ActionType.ABORT))
score_bad = env2.grade()
assert score_good > score_bad
def test_grader_deterministic():
"""Same episode, same score."""
def run_episode(seed):
env = RocketEnv(task="multi_fault_recovery", seed=seed)
env.reset()
for _ in range(5):
if env._done:
break
env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
return env.grade()
score1 = run_episode(99)
score2 = run_episode(99)
assert score1 == score2
# ─────────────────────────────────────────────────────────────────────────────
# Full episode tests
# ─────────────────────────────────────────────────────────────────────────────
def test_full_episode_nominal_launch():
"""Simulate a smart agent on nominal_launch."""
env = RocketEnv(task="nominal_launch", seed=7)
obs = env.reset()
rewards = []
for step in range(20):
if env._done:
break
# Smart strategy: diagnose faults, repair them, then authorize
faulted = None
for name, sub in env._state["subsystems"].items():
if sub.status != SubsystemStatus.NOMINAL:
faulted = name
break
if faulted and obs.step_number < 2:
action = RocketAction(action_type=ActionType.DIAGNOSE, target=faulted)
elif faulted:
action = RocketAction(action_type=ActionType.REPAIR, target=faulted)
else:
action = RocketAction(action_type=ActionType.AUTHORIZE)
result = env.step(action)
rewards.append(result.reward)
obs = result.observation
score = env.grade()
assert 0.0 <= score <= 1.0
assert len(rewards) > 0
def test_all_tasks_complete_without_error():
for task in TASK_CONFIGS:
env = RocketEnv(task=task, seed=42)
env.reset()
for _ in range(TASK_CONFIGS[task]["max_steps"]):
if env._done:
break
env.step(RocketAction(action_type=ActionType.RUN_DIAGNOSTIC))
score = env.grade()
assert 0.0 <= score <= 1.0
if __name__ == "__main__":
pytest.main([__file__, "-v"])