"""CPU-only adapter/reward checks in the selected mjlab Python environment.""" from types import SimpleNamespace import pytest torch = pytest.importorskip("torch") pytest.importorskip("mjlab") from universal_humanoid.config import Experiment # noqa: E402 from universal_humanoid.rewards import planar_velocity_exp # noqa: E402 from universal_humanoid.tasks import build_configs # noqa: E402 pytestmark = pytest.mark.integration @pytest.mark.parametrize("task", ["g1.flat", "t1.flat"]) def test_factories_are_independent(task): a, _ = build_configs(Experiment(task=task, reward_weights={"track_linear_velocity": 7})) b, _ = build_configs(Experiment(task=task)) assert a.rewards["track_linear_velocity"].weight == 7 assert b.rewards["track_linear_velocity"].weight != 7 @pytest.mark.parametrize( "exp, match", [ (Experiment(reward_weights={"typo": 1}), "Unknown reward weight"), (Experiment(reward_params={"track_linear_velocity": {"stdd": 1}}), "Unknown parameters"), ( Experiment(extra_rewards={"bad": {"function": "os.system", "weight": 1}}), "Unknown local", ), ( Experiment(extra_rewards={"pose": {"function": "joint_speed_l2", "weight": 1}}), "shadows", ), ], ) def test_reward_changes_fail_loudly(exp, match): with pytest.raises(ValueError, match=match): build_configs(exp) def test_evaluation_keeps_fall_terminations_and_finite_episodes(): exp = Experiment(episode_seconds=5) cfg, _ = build_configs(exp, evaluation=True, command=(0.3, 0.1, 0.2)) assert cfg.episode_length_s == 5 assert "fell_over" in cfg.terminations and "time_out" in cfg.terminations assert not cfg.curriculum assert all(event.mode == "reset" for event in cfg.events.values()) assert cfg.commands["twist"].ranges.lin_vel_x == (0.3, 0.3) assert cfg.commands["twist"].ranges.heading is None assert not cfg.commands["twist"].heading_command assert not cfg.observations["actor"].enable_corruption @pytest.mark.parametrize("task", ["g1.flat", "t1.flat"]) @pytest.mark.parametrize("profile", ["nominal", "randomized", "stress"]) def test_evaluation_randomization_profiles(task, profile): cfg, _ = build_configs(Experiment(task=task), evaluation=True, profile=profile) position = cfg.events["reset_robot_joints"].params["position_range"] assert position == ((0.0, 0.0) if profile == "nominal" else (-0.03, 0.03)) assert ("evaluation_push" in cfg.events) == (profile == "stress") assert cfg.observations["actor"].enable_corruption == (profile == "stress") assert not cfg.observations["critic"].enable_corruption assert "fell_over" in cfg.terminations and "time_out" in cfg.terminations assert all(term.mode in ("reset", "interval") for term in cfg.events.values()) def test_reward_shape_finiteness_and_direction(): velocity = torch.tensor([[0.3, 0, 0], [0.9, 0, 0]], dtype=torch.float32) env = SimpleNamespace( scene={"robot": SimpleNamespace(data=SimpleNamespace(root_link_lin_vel_b=velocity))}, command_manager=SimpleNamespace(get_command=lambda _: torch.tensor([[0.3, 0, 0]] * 2)), ) reward = planar_velocity_exp(env, std=0.5) assert reward.shape == (2,) assert torch.isfinite(reward).all() assert reward[0] == 1 and reward[1] < reward[0] with pytest.raises(ValueError, match="positive"): planar_velocity_exp(env, std=0)