File size: 17,084 Bytes
6aa6c12 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 | """SIMPLE interface checks without Isaac, MuJoCo, torch, or historical VLA models."""
from __future__ import annotations
import ast
import importlib.util
import json
import subprocess
import sys
import tempfile
import unittest
from dataclasses import replace
from pathlib import Path
from types import ModuleType, SimpleNamespace
from unittest.mock import Mock, patch
from benchmarks.simBenchmarks.SIMPLE.bench_server.adapter import (
ACTION_NAMES, ACTION_SPEC, HORIZONS, SEED, TASKS, SimpleHumanoidBenchmarkAdapter,
amo_payload, decode_action, public_observation, validate_tick_result,
)
from benchkit.errors import ProtocolError
class SimpleBoundaryTests(unittest.TestCase):
def test_import_has_no_native_or_model_dependencies(self):
code = "import sys; import benchmarks.simBenchmarks.SIMPLE.bench_server.adapter; "
code += "assert not any(n in sys.modules for n in ('torch','simple','mujoco','isaacsim','decoupled_wbc'))"
subprocess.run([sys.executable, "-c", code], check=True)
def test_named_36_mapping_preserves_legacy_left_hand_order(self):
command = decode_action(list(range(36)))
joints = command["joint_positions"]
self.assertEqual(len(set(ACTION_NAMES)), 36)
self.assertEqual(joints["left_hand_middle_0_joint"], 3)
self.assertEqual(joints["left_hand_index_0_joint"], 5)
self.assertEqual(joints["right_hand_index_0_joint"], 10)
self.assertEqual(joints["left_shoulder_pitch_joint"], 14)
self.assertEqual(joints["right_wrist_yaw_joint"], 27)
self.assertEqual(joints["waist_yaw_joint"], 30)
self.assertEqual(command["navigate_cmd"], (32, 33, 34, 35))
self.assertEqual(command["base_height"], 31)
def test_amo_waist_is_separate_from_fixed_native_torso(self):
payload = amo_payload(list(range(36)), (0.11, -0.15, 0.03))
self.assertEqual(payload["action_command"], [32, 35, 33, 30.25, 0.11, -0.15, 0.03, 34])
self.assertEqual(payload["waist_qpos"]["waist_roll_joint"], 28)
self.assertFalse(any(name.startswith("waist_") for name in payload["target_qpos"]))
self.assertEqual(len(payload["target_qpos"]), 28)
def test_actions_reject_wrong_length_nonfinite_bool_and_strings(self):
for vector in ([0] * 7, [0] * 35, [0] * 37,
[float("nan")] + [0] * 35, [float("inf")] + [0] * 35,
[True] + [0] * 35, ["0"] + [0] * 35):
with self.subTest(vector=vector), self.assertRaises((ValueError, TypeError)):
ACTION_SPEC.validate(vector)
def test_rgb_only_allowlist_and_copy(self):
import numpy as np
frame = np.ones((360, 640, 3), dtype=np.uint8)
raw = {"head_stereo_left": frame, "head_stereo_right": frame,
"joint_qpos": [7] * 43, "proprio": {"object_pose": [9]},
"success": True, "reward": 1, "environment_config": "secret", "distance": 0}
obs = public_observation(raw, 0, "native instruction")
self.assertEqual(tuple(obs.images), ("head_stereo_left",))
self.assertIsNone(obs.state)
self.assertEqual(obs.metadata, {})
frame[:] = 42
self.assertEqual(int(obs.images["head_stereo_left"][0, 0, 0]), 1)
for malformed in (np.zeros((360, 640, 4), dtype=np.uint8),
np.zeros((224, 224, 3), dtype=np.uint8), np.zeros((360, 640, 3), dtype=float)):
with self.assertRaises(ProtocolError):
public_observation({"head_stereo_left": malformed}, 0, "instruction")
def test_first_two_are_predetermined_not_success_selected(self):
episodes = tuple(SimpleHumanoidBenchmarkAdapter(task_names=(task,)).episodes(SEED)[0] for task in TASKS)
adapter = SimpleHumanoidBenchmarkAdapter()
self.assertEqual(tuple(item.task for item in episodes), TASKS)
self.assertEqual([item.metadata["reference_native_horizon"] for item in episodes], [400, 1000])
self.assertTrue(all(item.metadata["dr_level"] == 0 and item.metadata["episode_index"] == 0 for item in episodes))
with self.assertRaises(ProtocolError):
adapter.episodes(SEED + 1)
with self.assertRaises(ProtocolError):
adapter.open_episode(replace(episodes[0], metadata={**episodes[0].metadata, "episode_index": 1}))
with self.assertRaises(ValueError):
SimpleHumanoidBenchmarkAdapter(task_names=(TASKS[0], TASKS[0]))
with self.assertRaises(ValueError):
SimpleHumanoidBenchmarkAdapter(task_names=TASKS)
def test_backend_specific_names_and_distinct_sensor_budget_tracks(self):
mp = SimpleHumanoidBenchmarkAdapter()
sonic = SimpleHumanoidBenchmarkAdapter(task_names=(TASKS[1],))
self.assertEqual(mp.action_spec.names[34], "turning_flag")
self.assertEqual(sonic.action_spec.names[34], "turning_gate")
self.assertEqual(mp.observation_spec.state_names, ACTION_NAMES[:31])
self.assertEqual(mp.environment_step_cap, 1000)
self.assertEqual(mp.native_horizon, 400)
self.assertEqual(mp.budget_mode, "decisions")
self.assertEqual((mp.max_policy_calls, mp.replan_interval), (200, 5))
rgb = SimpleHumanoidBenchmarkAdapter(proprioception=False)
old = SimpleHumanoidBenchmarkAdapter(proprioception=False, budget_mode="horizon")
self.assertFalse(rgb.observation_spec.state_names)
self.assertEqual(old.environment_step_cap, 400)
self.assertEqual(len({a.descriptor.protocol_version for a in (mp, sonic, rgb, old)}), 4)
def test_qpos31_track_only_exposes_measured_named_robot_joints(self):
import numpy as np
raw = {"head_stereo_left": np.zeros((360, 640, 3), dtype=np.uint8),
"joint_positions": {name: index * .01 for index, name in enumerate(ACTION_NAMES[:31])},
"base_world_pose": [42] * 7, "object_qpos": [99], "amo_obs_command": [7] * 9}
obs = public_observation(raw, 1, "native instruction", True)
self.assertEqual(obs.state, tuple(index * .01 for index in range(31)))
self.assertEqual(obs.metadata, {})
self.assertIsNone(public_observation(raw, 1, "native instruction", False).state)
raw["joint_positions"]["object_joint"] = 0
with self.assertRaises(ProtocolError):
public_observation(raw, 1, "native instruction", True)
del raw["joint_positions"]["object_joint"]
raw["joint_positions"][ACTION_NAMES[0]] = float("nan")
with self.assertRaises(ProtocolError):
public_observation(raw, 1, "native instruction", True)
def test_decision_time_limit_does_not_end_at_reference_400(self):
import gymnasium as gym
class FakeNative(gym.Env):
metadata = {"max_episode_steps": 400}
def reset(self, **kwargs):
return {}, {}
def step(self, action):
return {}, 0., False, False, {}
raw = FakeNative()
env = gym.wrappers.TimeLimit(raw, max_episode_steps=SimpleHumanoidBenchmarkAdapter().environment_step_cap)
env.reset()
for _ in range(400):
result = env.step(None)
self.assertFalse(result[3])
for _ in range(600):
result = env.step(None)
self.assertTrue(result[3])
self.assertEqual(raw.metadata["max_episode_steps"], 400)
def test_diagnostic_hold_not_available_to_normal_session(self):
adapter = SimpleHumanoidBenchmarkAdapter()
session = adapter.open_episode(adapter.episodes(SEED)[0])
with self.assertRaises(ProtocolError):
session.diagnostic_hold()
session.close()
def test_rpc_rejects_truthiness_nonfinite_and_counter_errors(self):
valid = {"success": False, "terminated": False, "truncated": False, "reward": 0., "native_steps": 1}
self.assertEqual(validate_tick_result(valid, 1), (False, False, False, 0.))
for field in ("success", "terminated", "truncated"):
for value in ("false", "true", None, 0, 1, []):
with self.subTest(field=field, value=value), self.assertRaises(ProtocolError):
validate_tick_result({**valid, field: value}, 1)
for value in ("0", None, True, float("nan"), float("inf")):
with self.assertRaises(ProtocolError):
validate_tick_result({**valid, "reward": value}, 1)
for value in (True, 0, 2, "1", 1.0):
with self.assertRaises(ProtocolError):
validate_tick_result({**valid, "native_steps": value}, 1)
def test_private_success_info_never_enters_observation(self):
from benchkit.contracts import Observation
for diagnostic in (False, True):
adapter = SimpleHumanoidBenchmarkAdapter(diagnostic=diagnostic)
session = adapter.open_episode(adapter.episodes(SEED)[0])
session._rpc = Mock(return_value={"success": True, "terminated": True,
"truncated": False, "reward": .9, "native_steps": 1})
session._observation = Mock(return_value=Observation(1, "task"))
transition = session.diagnostic_hold() if diagnostic else session.step([0.] * 36)
self.assertEqual(transition.info, {"native_success": True, "native_steps": 1})
self.assertEqual(transition.observation.metadata, {})
self.assertIsNone(transition.observation.state)
def test_worker_step_calls_native_step_exactly_once_and_never_scores(self):
worker_file = Path(__file__).parents[1] / "environment/native_worker.py"
spec = importlib.util.spec_from_file_location("simple_private_worker_test", worker_file)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
with tempfile.TemporaryDirectory() as directory:
runtime = module.NativeRuntime()
runtime.artifact_dir = Path(directory)
runtime.raw = SimpleNamespace(_success=True)
runtime.env = SimpleNamespace(step=Mock(return_value=({}, .7, True, False, {"oracle": True})))
runtime._native_action = Mock(return_value="encoded-native-action")
runtime.capture = Mock(return_value={"frame": "rgb/private.png", "instruction": "task"})
result = runtime.step([0] * 36)
runtime.env.step.assert_called_once_with("encoded-native-action")
self.assertEqual(runtime.steps, 1)
self.assertTrue(result["success"])
self.assertNotIn("oracle", result)
with self.assertRaises(ValueError):
runtime.step([0] * 36)
tree = ast.parse(worker_file.read_text())
calls = [node.func.attr for node in ast.walk(tree) if isinstance(node, ast.Call) and isinstance(node.func, ast.Attribute)]
self.assertNotIn("compute_reward", calls)
self.assertNotIn("check_success", calls)
self.assertNotIn("load_dataset", calls)
def test_v2_passes_unclipped_finite_targets_to_native_amo(self):
worker_file = Path(__file__).parents[1] / "environment/native_worker.py"
spec = importlib.util.spec_from_file_location("simple_native_target_test", worker_file)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
action_module = ModuleType("simple.core.action")
action_module.ActionCmd = lambda kind, **parameters: SimpleNamespace(type=kind, parameters=parameters)
runtime = module.NativeRuntime()
runtime.agent = None
runtime.amo_torso_ypr = (0., -.15, 0.)
# No joint-limit capability is needed to convert a public command.
runtime.robot = SimpleNamespace()
vector = [0.] * 36
vector[3] = 1.5208409952754696e-06 # public measured value from the closed failed probe
vector[14] = 4.0 # finite target outside the MJCF's physical joint range
vector[31] = .75
with patch.dict(sys.modules, {"simple.core.action": action_module}):
converted = runtime._native_action(vector)
self.assertEqual(converted.type, "eval_move_actuators")
self.assertEqual(converted.parameters["target_qpos"][ACTION_NAMES[3]], vector[3])
self.assertEqual(converted.parameters["target_qpos"][ACTION_NAMES[14]], 4.)
self.assertEqual(converted.parameters["action_command"], [0., 0., 0., 0., 0., -.15, 0., 0.])
for invalid in ([0.] * 35, [float("nan")] + vector[1:], [True] + vector[1:]):
with self.assertRaises((ValueError, TypeError)):
runtime._native_action(invalid)
self.assertEqual(vector[3], 1.5208409952754696e-06)
def test_v2_declares_no_added_command_bounds_and_keeps_native_shape(self):
for task in TASKS:
adapter = SimpleHumanoidBenchmarkAdapter(task_names=(task,))
self.assertTrue(adapter.descriptor.protocol_version.endswith("/v2"))
self.assertIsNone(adapter.action_spec.low)
self.assertIsNone(adapter.action_spec.high)
values = [0.] * 36
values[3] = 1.5208409952754696e-06
values[31] = .75
self.assertEqual(adapter.action_spec.validate(values), tuple(values))
self.assertEqual(adapter.action_spec.control_hz, 50.)
def test_service_malformed_rejection_preserves_sequence_then_accepts_raw_encoder(self):
import numpy as np
from benchkit.contracts import Observation, Transition
from benchkit.sim_server.server import BenchmarkService, ServiceActionRejected
adapter = SimpleHumanoidBenchmarkAdapter()
image = np.zeros((360, 640, 3), dtype=np.uint8)
def observation(step):
return Observation(step, "fixture", images={"head_stereo_left": image}, state=(0.,) * 31)
native = SimpleNamespace(reset=Mock(return_value=observation(0)),
step=Mock(return_value=Transition(observation(1), 0., False, False)),
metrics=lambda: {}, artifacts=lambda: {}, close=Mock())
with tempfile.TemporaryDirectory() as directory, patch.object(adapter, "open_episode", return_value=native):
service = BenchmarkService(adapter, root=Path(directory) / "service", seed=SEED)
try:
sid = service.dispatch("open", {"episode_id": adapter.episodes()[0].id})["session_id"]
service.dispatch("reset", {"session_id": sid, "sequence": 0})
with self.assertRaises(ServiceActionRejected):
service.dispatch("step", {"session_id": sid, "sequence": 1, "action": [0.] * 35})
native.step.assert_not_called()
self.assertEqual(service.sessions[sid]["sequence"], 1)
self.assertEqual(service.sessions[sid]["state"], "running")
action = [0.] * 36
action[3], action[31] = 1.5208409952754696e-06, .75
service.dispatch("step", {"session_id": sid, "sequence": 1, "action": action})
native.step.assert_called_once_with(tuple(action))
self.assertEqual(service.sessions[sid]["controls"], 1)
self.assertEqual(service.sessions[sid]["sequence"], 2)
finally:
service.close()
def test_initial_success_unstable_reset_and_source_drift_fail_closed(self):
worker_file = Path(__file__).parents[1] / "environment/native_worker.py"
spec = importlib.util.spec_from_file_location("simple_private_guard_test", worker_file)
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module)
module.check_initial_state(False)
module.check_initial_state(False, True)
for success, stable in ((True, None), (True, True), (False, False), ("false", True), (None, True)):
with self.assertRaises(ValueError):
module.check_initial_state(success, stable)
with tempfile.TemporaryDirectory() as directory:
path = Path(directory) / "native_source.py"
path.write_text("first")
expected = module.fingerprint(path)["sha256"]
module.require_hash(path, expected)
path.write_text("drifted")
with self.assertRaises(ValueError):
module.require_hash(path, expected)
def test_operator_parameters_and_episode_limits_are_strict(self):
from benchkit.registry import create_benchmark
for args in ({"gpu_id": True}, {"gpu_id": -1}, {"gpu_id": "4"},
{"worker_timeout": float("nan")}, {"worker_timeout": 0},
{"worker_timeout": True}, {"diagnostic": "false"}):
with self.subTest(args=args), self.assertRaises(ValueError):
create_benchmark("simple_humanoid", args)
adapter = create_benchmark("simple_humanoid")
for limit in (True, 0, -1, 1.5):
with self.assertRaises(ValueError):
adapter.episodes(limit=limit)
self.assertEqual(len(adapter.episodes(limit=1)), 1)
if __name__ == "__main__":
unittest.main()
|