Files
chenlin c32df70326
Checks / Layout, translation and Colab checks (push) Has been cancelled
Checks / Ruff (push) Has been cancelled
Checks / Framework tests (native backend, CPU) (push) Has been cancelled
Checks / Portable Rust controller (push) Has been cancelled
release: v0.1 generic robot training and application platform
2026-11-05 11:05:01 +08:00

108 lines
3.9 KiB
Python

"""The shared curriculum needs mature evidence on both commanded axes."""
from types import SimpleNamespace
import pytest
torch = pytest.importorskip("torch")
from mjrl.tasks.curriculum import CommandRangeCurriculum
PARAMS = {
"fractions": (0.4, 0.7, 1.0),
"ceilings": (1.0, 0.0, 0.5),
"command_name": "twist",
"dwell_steps": 150,
"min_steps": 10,
"min_samples": 3,
"gate_ratio": 0.4,
"ema_alpha": 1.0,
}
def _inputs(monkeypatch):
monkeypatch.delenv("MJRL_COMMAND_LEVEL", raising=False)
ranges = SimpleNamespace()
term = SimpleNamespace(
cfg=SimpleNamespace(ranges=ranges, resampling_time_range=(3.0, 7.0)),
metrics={"error_vel_xy": torch.zeros(3), "error_vel_yaw": torch.zeros(3)},
)
env = SimpleNamespace(
command_manager=SimpleNamespace(get_term=lambda _: term),
common_step_counter=0,
max_episode_length=120,
episode_length_buf=torch.tensor([70, 70, 70]),
step_dt=0.1,
)
ladder = CommandRangeCurriculum(SimpleNamespace(params=PARAMS), env)
return env, term, ladder
def test_curriculum_waits_for_real_rollouts_and_pools_staggered_resets(monkeypatch) -> None:
env, term, ladder = _inputs(monkeypatch)
report = ladder(env, torch.arange(3), **PARAMS)
assert report["level"] == 0
assert term.cfg.ranges.lin_vel_x == pytest.approx((-0.4, 0.4))
assert ladder.err_ema is None
env.common_step_counter = 119
ladder(env, torch.arange(3), **PARAMS)
assert ladder.err_ema is None, "initial randomized episode ages must not count"
env.common_step_counter = 151
for index in (0, 1):
report = ladder(env, torch.tensor([index]), **PARAMS)
assert report["level"] == 0
assert ladder.err_ema is None
report = ladder(env, torch.tensor([2]), **PARAMS)
assert report["level"] == 1
assert term.cfg.ranges.lin_vel_x == pytest.approx((-0.7, 0.7))
assert term.cfg.ranges.lin_vel_y == (0.0, 0.0)
assert term.cfg.ranges.ang_vel_z == pytest.approx((-0.35, 0.35))
assert ladder.err_ema is None, "promotion must discard evidence from the previous stage"
@pytest.mark.parametrize("blocked", ["error_vel_xy", "error_vel_yaw"])
def test_bad_tracking_on_either_axis_blocks_promotion(monkeypatch, blocked) -> None:
env, term, ladder = _inputs(monkeypatch)
env.common_step_counter = 200
term.metrics[blocked].fill_(0.3)
assert ladder(env, torch.arange(3), **PARAMS)["level"] == 0
term.metrics[blocked].zero_()
assert ladder(env, torch.arange(3), **PARAMS)["level"] == 1
@pytest.mark.parametrize("saved", [None, {"level": 0, "last_promotion_step": 1}])
def test_resume_ignores_randomized_episode_ages_until_a_fresh_horizon(monkeypatch, saved) -> None:
env, term, ladder = _inputs(monkeypatch)
start = 10_000
ladder.load_state_dict(saved, step=start)
assert "_evidence_start_step" not in ladder.state_dict()
for offset in (0, 1, env.max_episode_length - 1):
env.common_step_counter = start + offset
assert ladder(env, torch.arange(3), **PARAMS)["level"] == 0
assert ladder.err_ema is None
assert term.cfg.ranges.lin_vel_x == pytest.approx((-0.4, 0.4))
env.common_step_counter += 1
assert ladder(env, torch.arange(3), **PARAMS)["level"] == 1
def test_loading_again_restarts_the_horizon_and_drops_partial_samples(monkeypatch) -> None:
env, _, ladder = _inputs(monkeypatch)
start = 10_000
saved = ladder.state_dict()
ladder.load_state_dict(saved, step=start)
env.common_step_counter = start + env.max_episode_length
assert ladder(env, torch.tensor([0, 1]), **PARAMS)["level"] == 0
start = 20_000
ladder.load_state_dict(saved, step=start)
env.common_step_counter = start + env.max_episode_length - 1
assert ladder(env, torch.arange(3), **PARAMS)["level"] == 0
env.common_step_counter += 1
assert ladder(env, torch.tensor([2]), **PARAMS)["level"] == 0
assert ladder(env, torch.tensor([0, 1]), **PARAMS)["level"] == 1