File size: 17,911 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 | """Fixed-manifest DOMINO direct qpos service; native dynamics/scoring stay private."""
from __future__ import annotations
import gc
import hashlib
from pathlib import Path
from typing import Mapping
from benchkit.artifacts import append_jsonl, atomic_write_json
from benchkit.contracts import ActionSpec, BenchmarkDescriptor, EpisodeSpec, Observation, ObservationSpec, Transition
from benchkit.errors import ProtocolError
from benchkit.registry import register_benchmark
from benchmarks.simBenchmarks.DOMINO.bench_server.protocol import (
ACTION_NAMES, CAMERAS, IMAGE_SHAPE, INSTRUCTION, NATIVE_HORIZON,
NATIVE_REVISION, PROTOCOL, STATE_NAMES, TASK, TASK_CONFIG,
)
from benchmarks.simBenchmarks.DOMINO.environment import runtime
from benchmarks.simBenchmarks.DOMINO.environment.native_outcome import NativeOutcome
def policy_observation(raw, step: int) -> Observation:
import numpy as np
if not isinstance(raw, Mapping) or not isinstance(raw.get("observation"), Mapping):
raise ProtocolError("DOMINO observation must contain native cameras")
images = {}
for name in CAMERAS:
item = raw["observation"].get(name)
frame = item.get("rgb") if isinstance(item, Mapping) else None
if not isinstance(frame, np.ndarray) or frame.shape != IMAGE_SHAPE or frame.dtype != np.uint8:
raise ProtocolError("DOMINO cameras require native 240x320 uint8 RGB")
images[name] = np.array(frame, copy=True, order="C")
command = raw.get("joint_action")
values = command.get("vector") if isinstance(command, Mapping) else None
if not isinstance(values, (list, tuple, np.ndarray)) or np.shape(values) != (14,):
raise ProtocolError("DOMINO commanded targets require 14 values")
state = ActionSpec("commanded_targets", 14).validate(values)
# Explicit native commanded drive targets, NOT measured encoders. Never
# forward native camera calibration, object state, scores or motion payload.
return Observation(step=step, instruction=INSTRUCTION, images=images, state=state,
metadata={"state_provenance": "native robot commanded drive/gripper targets; not measured encoders"})
class DOMINOEpisode:
def __init__(self, benchmark, episode):
self.benchmark, self.episode = benchmark, episode
self.env = self.outcome = self.counter = self.cache = None
self.artifact_dir = None
self.writers = {}
self.reset_attempted = self.reset_completed = self.closed = self.finished = False
self.finalized = self.control_error = False
self.native_terminal = False
self.close_error = None
self.step_count = self.acknowledged_controls = self.native_steps = self.physics_ticks = 0
self.native_metrics = {"success": False, "manipulation_score": None, "route_completion": None}
def bind_artifact_dir(self, path):
if self.artifact_dir is not None or self.reset_attempted or self.closed:
raise ProtocolError("bind DOMINO artifacts once before reset")
path = Path(path)
path.mkdir(parents=True, exist_ok=True)
self.artifact_dir = path.resolve()
def _log(self, event, **values):
if self.artifact_dir is not None:
append_jsonl(self.artifact_dir / "native_events.jsonl", {"event": event, **values})
def reset(self):
if self.reset_attempted or self.closed or self.finalized:
raise ProtocolError("DOMINO permits one reset attempt")
self.reset_attempted = True
try:
if self.artifact_dir is None:
raise ProtocolError("DOMINO requires bound artifacts and an owned JIT cache")
self.cache = self.artifact_dir / "cache/curobo-jit"
self.cache.mkdir(parents=True, exist_ok=False)
benchmark = self.benchmark
provenance = runtime.readiness(benchmark.source_root, benchmark.deployment_manifest_sha256)
atomic_write_json(self.artifact_dir / "native_provenance.json", provenance)
with runtime.native_directory(benchmark.source_root, self.cache):
config = runtime.resolved_config(benchmark.source_root)
task_class, tracker_class, manifest_class = runtime.import_native(benchmark.source_root)
entry = runtime.load_entry(benchmark.episode_manifest, benchmark.episode_manifest_sha256,
benchmark.native_seed, config, manifest_class)
self.env = task_class()
self._log("reset_start", native_seed=benchmark.native_seed, expert_called=False)
self.env.setup_demo(now_ep_num=0, seed=benchmark.native_seed, is_test=True, **config)
self.env._saved_dynamic_motion_info = entry["dynamic_motion_info"]
self.env.set_instruction(INSTRUCTION)
if not self.env.init_dynamic_motion_for_eval():
raise ProtocolError("fixed manifest dynamic initialization rejected; no seed replacement")
compiled = runtime.verify_compiled(self.env)
self.outcome = NativeOutcome(self.env, config, tracker_class)
self.counter = runtime.SceneStepCounter(self.env.scene)
self.env.scene = self.counter
provenance.update({"protocol_version": PROTOCOL, "compiled_contract": compiled,
"episode_manifest_sha256": benchmark.episode_manifest_sha256,
"native_seed": benchmark.native_seed, "expert_called": False,
"instruction_selection": "fixed native full_description, not randomized episode wording",
"state_provenance": "native commanded targets; not encoders",
"initialization_physics_ticks": None,
"video_timebase": "10fps presentation, not physical control frequency"})
observation = self._observe()
provenance["loaded_oidn_files"] = runtime.verify_loaded_renderer(provenance)
atomic_write_json(self.artifact_dir / "native_provenance.json", provenance)
self.native_metrics = self.outcome.metrics(benchmark.native_seed)
self.reset_completed = True
self._log("reset_end", native_controls=0, initialization_physics_ticks=None)
return observation
except BaseException:
self.finished = True
try:
self.close()
except BaseException as exc:
self._log("reset_cleanup_error", error_type=type(exc).__name__)
raise
def _observe(self):
observation = policy_observation(self.env.get_obs(), self.step_count)
if self.artifact_dir is not None:
import imageio.v2 as imageio
from PIL import Image
for camera, frame in observation.images.items():
path = self.artifact_dir / f"{camera}-{self.step_count:06d}.png"
Image.fromarray(frame).save(path)
if camera not in self.writers:
self.writers[camera] = imageio.get_writer(str(self.artifact_dir / f"{camera}.mp4"),
fps=10, codec="libx264", macro_block_size=1)
self.writers[camera].append_data(frame)
append_jsonl(self.artifact_dir / "frames.jsonl", {
"camera": camera, "step": self.step_count, "path": path.name,
"sha256": hashlib.sha256(path.read_bytes()).hexdigest(),
"rgb_sha256": hashlib.sha256(frame.tobytes()).hexdigest(),
"policy_physics_ticks": self.physics_ticks})
return observation
def step(self, action):
if not self.reset_completed or self.closed or self.finished or self.finalized:
raise ProtocolError("DOMINO control requires a live initialized episode")
values = self.benchmark.action_spec.validate(action)
if self.native_steps >= NATIVE_HORIZON:
raise ProtocolError("DOMINO horizon exhausted")
self._log("action_start", action=list(values), control=self.step_count + 1)
try:
with runtime.native_directory(self.benchmark.source_root, self.cache):
self.env.take_action(values, action_type="qpos")
self.native_steps = int(self.env.take_action_cnt)
self.physics_ticks = self.counter.steps
if self.native_steps != self.step_count + 1:
raise ProtocolError("native DOMINO target count differs")
self.step_count += 1
# Native take_action has already called tracker.on_step once.
self.outcome.after_action()
self.finished = self.outcome.finished
self.native_terminal = self.finished
self.native_metrics = self.outcome.metrics(self.benchmark.native_seed)
if self.native_metrics["metrics_tracker_steps"] != self.native_steps:
raise ProtocolError("native metrics tracker count differs from completed controls")
observation = self._observe()
self._log("action_end", native_controls=self.native_steps,
policy_physics_ticks=self.physics_ticks, outcome=self.native_metrics)
result = Transition(observation, reward=float(self.outcome.success),
terminated=self.finished and (self.outcome.success or self.outcome.fail_reason is not None),
truncated=self.finished and not self.outcome.success and self.outcome.fail_reason is None,
info={"native_success": self.outcome.success, "native_steps": self.native_steps})
self.acknowledged_controls += 1
return result
except BaseException:
self.control_error = self.finished = True
self.native_steps = int(getattr(self.env, "take_action_cnt", self.native_steps))
self.physics_ticks = self.counter.steps
self._log("action_error", native_started_controls=self.native_steps,
acknowledged_controls=self.acknowledged_controls,
completed_physics_ticks=self.physics_ticks, replay_allowed=False)
raise
def metrics(self):
return {**self.native_metrics, "native_horizon": NATIVE_HORIZON,
"native_steps": self.native_steps, "completed_native_controls": self.step_count,
"acknowledged_controls": self.acknowledged_controls,
"policy_physics_ticks": self.physics_ticks, "policy_simulated_seconds": self.physics_ticks / 250,
"reset_attempted": self.reset_attempted, "reset_completed": self.reset_completed,
"initialization_physics_ticks": None, "control_error": self.control_error,
"native_terminal": self.native_terminal,
"native_score_complete": self.native_terminal and not self.control_error,
"score_scope": "terminal native outcome" if self.native_terminal and not self.control_error else "partial diagnostic/stop/error state; not a completed native task score",
"native_source_revision": NATIVE_REVISION,
"reward_semantics": "binary native success; SR is not MS/100"}
def _close_writers(self):
writers, self.writers = self.writers, {}
errors = []
for writer in writers.values():
try:
writer.close()
except BaseException as exc:
errors.append(exc)
if errors:
raise ProtocolError("DOMINO media finalization failed") from errors[0]
def artifacts(self):
self.finalized = True
self._close_writers()
if self.artifact_dir is None:
return {}
atomic_write_json(self.artifact_dir / "native_result.json", self.metrics())
files = {camera + "_video": camera + ".mp4" for camera in CAMERAS}
files.update({key: key + ".json" for key in ("native_result", "native_provenance", "native_cleanup")})
files.update(frames="frames.jsonl", native_events="native_events.jsonl")
return {key: str(self.artifact_dir / filename) for key, filename in files.items()
if (self.artifact_dir / filename).is_file()}
def close(self):
if self.closed:
if self.close_error:
raise ProtocolError("previous DOMINO cleanup failed") from self.close_error
return
self.closed = True
errors = []
try:
self._close_writers()
except BaseException as exc:
errors.append(exc)
env, self.env = self.env, None
closed = released = False
if env is not None:
if self.counter is not None:
env.scene = self.counter.scene
with runtime.native_directory(self.benchmark.source_root, self.cache):
try:
env._metrics_tracker = None
env.close_env()
closed = True
except BaseException as exc:
errors.append(exc)
try:
env.release_episode_resources()
released = True
except BaseException as exc:
errors.append(exc)
self.counter = self.outcome = None
del env
gc.collect()
if self.artifact_dir is not None:
try:
atomic_write_json(self.artifact_dir / "native_cleanup.json", {
"native_close_returned": closed, "release_resources_returned": released,
"error_types": [type(e).__name__ for e in errors],
"scope": "native close/release; process exit separately verifies final cleanup"})
except BaseException as exc:
errors.append(exc)
if errors:
self.close_error = errors[0]
raise ProtocolError("DOMINO cleanup failed") from errors[0]
class DOMINOBenchmarkAdapter:
budget_mode = "horizon"
required_replan_interval = 1
replan_interval = 1
environment_step_cap = NATIVE_HORIZON
def __init__(self, source_root, deployment_manifest_sha256, episode_manifest,
episode_manifest_sha256, native_seed=100000, task=TASK, evaluation_mode="fast"):
if task != TASK or evaluation_mode != "fast":
raise ValueError("DOMINO v1 supports only one fixed click_bell fast case, not full")
for value in (source_root, episode_manifest):
if not isinstance(value, str) or not Path(value).is_absolute() or ".." in Path(value).parts:
raise ValueError("DOMINO source and trusted manifest require explicit absolute paths")
for value in (deployment_manifest_sha256, episode_manifest_sha256):
if not isinstance(value, str) or len(value) != 64 or any(c not in "0123456789abcdef" for c in value):
raise ValueError("DOMINO requires SHA256-pinned deployment and trusted manifest")
if type(native_seed) is not int or not 0 <= native_seed < 2**32:
raise ValueError("native_seed must be a uint32 integer")
self.source_root, self.episode_manifest = Path(source_root), Path(episode_manifest)
self.deployment_manifest_sha256 = deployment_manifest_sha256
self.episode_manifest_sha256, self.native_seed = episode_manifest_sha256, native_seed
self.claimed = False
@property
def descriptor(self):
return BenchmarkDescriptor("domino", "9c94f2d-fixed-manifest-sm120", PROTOCOL, NATIVE_REVISION)
@property
def observation_spec(self):
return ObservationSpec(cameras=CAMERAS, state_names=STATE_NAMES, modalities=("language", "rgb", "state"))
@property
def action_spec(self):
return ActionSpec("domino_aloha_absolute_joint_targets", 14, ACTION_NAMES,
frame="native absolute Aloha arm joint coordinates in radians", control_hz=None,
rotation="joint coordinates; no EEF or task planner RPC",
gripper="native normalized absolute command, nominal 0 closed/1 open; native mapping/clipping unchanged")
def episodes(self, seed, limit=None):
if type(seed) is not int or not 0 <= seed < 2**32:
raise ValueError("harness seed must be uint32")
if limit is not None and (type(limit) is not int or limit != 1):
raise ValueError("DOMINO fixed task service contains exactly one selected case")
return (EpisodeSpec(f"{TASK}-native-seed{self.native_seed}-episode000", TASK, seed, INSTRUCTION,
{"native_seed": self.native_seed, "native_horizon": NATIVE_HORIZON,
"task_config": TASK_CONFIG, "evaluation_mode": "fast",
"episode_manifest_sha256": self.episode_manifest_sha256,
"sample_selection": "one predeclared private fixed-manifest entry; no online expert or retry"}),)
def open_episode(self, episode):
if episode not in self.episodes(episode.seed) or self.claimed:
raise ProtocolError("DOMINO case mismatched or already claimed")
self.claimed = True
return DOMINOEpisode(self, episode)
@register_benchmark("domino")
def make_domino(parameters):
allowed = {"source_root", "deployment_manifest_sha256", "episode_manifest",
"episode_manifest_sha256", "native_seed", "task", "evaluation_mode"}
if set(parameters) - allowed:
raise ValueError("unknown DOMINO server parameters")
return DOMINOBenchmarkAdapter(**parameters)
|