feat(envs): lazy env init + AsyncVectorEnv as default for n_envs > 1 (#3274)

* docs(benchmarks): add benchmark integration guide and standardize benchmark docs Add a comprehensive guide for adding new benchmarks to LeRobot, and refactor the existing LIBERO and Meta-World docs to follow the new standardized template. Made-with: Cursor * refactor(envs): move dispatch logic from factory into EnvConfig subclasses Replace hardcoded if/elif chains in factory.py with create_envs() and get_env_processors() methods on EnvConfig. New benchmarks now only need to register a config subclass — no factory.py edits required. Net -23 lines: factory.py shrinks from ~200 to ~70 lines of logic. Made-with: Cursor * docs(benchmarks): clean up adding-benchmarks guide for clarity Rewrite for simpler language, better structure, and easier navigation. Move quick-reference table to the top, fold eval explanation into architecture section, condense the doc template to a bulleted outline. Made-with: Cursor * fix link * fix task count * fix: enable SmolVLA eval on LIBERO with custom camera mappings - Thread camera_name_mapping from LiberoEnv config through to gym envs - Sync features_map with camera_name_mapping in LiberoEnv.__post_init__ - Fix render() to use first available camera instead of hardcoded "image" - Handle non-dict final_info in rollout by falling back to info["is_success"] - Add use_peft legacy field to SmolVLAConfig for checkpoint compat - Add defaults to GR00TN15Config init=False fields for transformers 5.3 Made-with: Cursor * fix: use direct AutoresetMode import for gymnasium compat Made-with: Cursor * fix: handle gymnasium < 1.0 without AutoresetMode Made-with: Cursor * refactor: revert policy changes, keep env-only camera mapping fixes - Revert GR00T N1.5 default_factory/default changes (transformers compat) - Revert SmolVLA use_peft legacy field - Apply ruff formatting fixes - camera_name_mapping stays entirely in env/eval layer (no policy changes) Made-with: Cursor * Update docs/source/env_processor.mdx Co-authored-by: Khalil Meftah <khalil.meftah@huggingface.co> Signed-off-by: Pepijn <138571049+pkooij@users.noreply.github.com> * feat(envs): lazy env init + AsyncVectorEnv as default for n_envs > 1 LiberoEnv and MetaworldEnv previously allocated GPU resources (EGL context, OpenGL framebuffer) in __init__, before AsyncVectorEnv's fork(). Worker processes inherited stale GPU handles, causing EGL_BAD_CONTEXT crashes on first render. Fix: defer OffScreenRenderEnv / MT1 construction to _ensure_env(), called on first reset() or step() inside the worker subprocess. Each worker creates its own clean context after fork(). Also fixes lerobot_eval.py:170 (add_envs_task TODO): replace with env.call("task") which works with both SyncVectorEnv and AsyncVectorEnv. AsyncVectorEnv is now the default for n_envs > 1; auto-downgraded to SyncVectorEnv when n_envs=1 (no benefit, less overhead). Expected speedup: ~15-20x for LIBERO Spatial with batch_size=50. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: close envs between tasks to prevent worker process accumulation eval_policy_all never closed environments after each task completed, causing AsyncVectorEnv worker processes to accumulate (N_tasks × n_envs). This led to OOM, BrokenPipeError and EOFError on multi-task benchmarks. Also fixes: - AsyncVectorEnv compat in envs/utils.py (use get_attr/call instead of .envs) - Tuple task handling in tokenizer_processor and lerobot_eval - _LazyAsyncVectorEnv for deferred worker spawning in LIBERO Made-with: Cursor * fix(eval): use task_description instead of task for language conditioning env.call("task") returns the LIBERO task name with underscores (e.g. "pick_up_the_black_bowl_...") instead of the natural language description ("pick up the black bowl ..."). The VLM tokenizes these completely differently, causing 0.0 reward across all episodes. Made-with: Cursor * docs: update adding_benchmarks for async env changes - Replace add_envs_task reference with env.call("task_description") - Update use_async_envs default to True - Add note about lazy GPU init for AsyncVectorEnv compatibility Made-with: Cursor * feat(eval): batch_size=auto + faster env loading - batch_size=0 (default) auto-tunes based on CPU cores, capped by n_episodes and 64. Removes the need for users to guess the right value. The old batch_size > n_episodes error is replaced by silently clamping to n_episodes. - _LazyAsyncVectorEnv accepts pre-computed spaces so only one temp env is created per suite (not per task). For libero_spatial (10 tasks) this avoids 9 redundant LiberoEnv instantiations during env setup. Made-with: Cursor * docs: add evaluation guide and update benchmarks doc - New docs/source/evaluation.mdx covering lerobot-eval usage, batch_size auto-tuning, AsyncVectorEnv performance, tuning tips, output format, multi-task evaluation, and programmatic usage. - Add evaluation page to _toctree.yml under Benchmarks section. - Update adding_benchmarks.mdx to reference batch_size auto default and link to the evaluation guide. Made-with: Cursor * docs(evaluation): remove benchmark table, rename section header Made-with: Cursor * perf(eval): shared memory, observation passthrough, task prefetch - AsyncVectorEnv now uses shared_memory=True for zero-copy observation transfer - LiberoEnvConfig.gym_kwargs passes observation_height/width to the env - eval_policy_all prefetches next task's workers while current task runs Made-with: Cursor * style: ruff format Made-with: Cursor * chore: revert env_processor.mdx changes (not part of this PR) Made-with: Cursor * ci(benchmarks): add isolated integration tests for libero and metaworld Each benchmark gets its own Docker image (lerobot[libero] / lerobot[metaworld] only) so incompatible dep trees cannot collide. A 1-episode smoke eval runs per benchmark on GPU runners. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * ci(benchmarks): pin action hashes and use uv sync --locked Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * ci(benchmarks): trigger only on envs/ or lerobot_eval.py changes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(ci): set LIBERO_DATA_FOLDER to bypass interactive stdin prompt libero/__init__.py calls input() to ask about a custom dataset path, which raises EOFError when stdin is closed inside Docker. Setting LIBERO_DATA_FOLDER skips the prompt entirely. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * docs(benchmarks): add CI smoke test step to adding_benchmarks guide Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(ci): pre-create libero config in Dockerfile to bypass stdin prompt libero/__init__.py calls input() when ~/.libero/config.yaml is missing. We write the config at image build time (without importing libero) so the prompt never fires at runtime. Also trigger CI on pyproject.toml changes. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(ci): use shell to create libero config instead of multiline python -c The multiline RUN python -c "..." was being parsed as Dockerfile instructions. Use printf to write ~/.libero/config.yaml directly. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(ci): point libero config to bundled package init_files The config was pointing to /tmp/libero_init which doesn't exist. Use importlib.util.find_spec to locate the hf-libero package directory and write paths to the actual bundled bddl_files/init_files/assets. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(ci): add smolvla extra to benchmark Dockerfiles num2words (required by SmolVLM processor) is declared in lerobot[smolvla], not lerobot[libero/metaworld]. Install both extras together. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(eval): render_frame covers _LazyAsyncVectorEnv isinstance(env, AsyncVectorEnv) silently skipped _LazyAsyncVectorEnv, causing video rendering to produce no frames on the default async path. Switch to hasattr(env, "call") so any async-compatible env (including _LazyAsyncVectorEnv) hits the call("render") branch. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * refactor(envs): remove unused _get_sub_env_attr helper _get_sub_env_attr was defined but never called anywhere in the codebase. _sub_env_has_attr (its sibling) is kept — it is actively used in utils.py. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * chore: apply prettier formatting to docs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * docs(env_processor): remove deprecated add_envs_task from pipeline example add_envs_task is replaced by env.call("task_description") in this PR. Remove it from the pipeline walkthrough and renumber the steps (8→7). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * refactor(envs): remove __del__ from _LazyAsyncVectorEnv __del__ is unreliable as a cleanup mechanism. close() is already called explicitly in the eval loop's finally block, so the finalizer is redundant. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(eval): prefetch next task's workers after close to avoid GPU memory overlap Previously, next task's AsyncVectorEnv workers were spawned while the current task was still running, causing both tasks' GPU contexts to coexist. Moving the prefetch start into the finally block (after env.close()) ensures workers for task N+1 only spin up once task N has released GPU memory. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * refactor(envs): move _LazyAsyncVectorEnv to utils and apply to metaworld _LazyAsyncVectorEnv lived in libero.py but metaworld had the same OOM problem: all tasks' AsyncVectorEnv workers were spawned eagerly, wasting GPU memory for tasks not yet running. Move the class to envs/utils.py so both environments share it, then apply the same is_async + lazy wrapping pattern in create_metaworld_envs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * chore: remove out-of-scope benchmark/CI/docs files from PR Benchmark CI workflow, Dockerfiles, benchmark docs, evaluation smoke-test doc, and dispatch tests belong in a separate PR. Scope this PR to the async env init changes only. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * chore: restore adding_benchmarks + test_dispatch, drop env_processor changes - Restore docs/source/adding_benchmarks.mdx (belongs in this PR) - Restore tests/envs/test_dispatch.py (belongs in this PR) - Revert docs/source/env_processor.mdx to main (out of scope for this PR) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * docs(adding_benchmarks): remove CI smoke test step (coming in separate PR) Step 7 (Dockerfile + benchmark_tests.yml CI job) and its table rows are out of scope for this PR. The CI infrastructure will be added on top in a follow-up PR. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * refactor(envs): remove unused add_envs_task Replaced by env.call("task_description") in lerobot_eval.py. No callers remain in the codebase. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * style: fix prettier formatting in env_processor.mdx Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(eval): catch AttributeError and NotImplementedError explicitly for task description Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(envs): use forkserver context and close envs in test to prevent deadlock AsyncVectorEnv with default fork context leaks worker processes between test_policy parametrized cases; subsequent env creation deadlocks because new forked workers inherit stale pipe FDs from previous test's leaked workers. - configs.py: pass context="forkserver" to AsyncVectorEnv (matches _LazyAsyncVectorEnv) - test_policies.py: call close_envs(envs) at end of test_policy to clean up workers Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix(envs): default use_async_envs=False in create_envs and make_env Tests that call make_env(n_envs=2) without passing use_async_envs were getting AsyncVectorEnv, whose forked workers can't resolve gym namespaces registered at runtime. Default to False (sync) so existing tests pass. lerobot_eval.py explicitly passes cfg.eval.use_async_envs, so the CLI async behaviour (controlled by EvalConfig.use_async_envs) is unchanged. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Signed-off-by: Pepijn <138571049+pkooij@users.noreply.github.com> Co-authored-by: Khalil Meftah <khalil.meftah@huggingface.co> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-05-16 00:59:46 +00:00 · 2026-04-09 10:29:20 +02:00
parent 5de7aa5a4f
commit 919184d6f8
13 changed files with 331 additions and 159 deletions
@@ -65,20 +65,27 @@ class WandBConfig:
 class EvalConfig:
    n_episodes: int = 50
    # `batch_size` specifies the number of environments to use in a gym.vector.VectorEnv.
-    batch_size: int = 50
+    # Set to 0 for auto-tuning based on available CPU cores and n_episodes.
+    batch_size: int = 0
    # `use_async_envs` specifies whether to use asynchronous environments (multiprocessing).
-    use_async_envs: bool = False
+    # Defaults to True; automatically downgraded to SyncVectorEnv when batch_size=1.
+    use_async_envs: bool = True

    def __post_init__(self) -> None:
+        if self.batch_size == 0:
+            self.batch_size = self._auto_batch_size()
        if self.batch_size > self.n_episodes:
-            raise ValueError(
-                "The eval batch size is greater than the number of eval episodes "
-                f"({self.batch_size} > {self.n_episodes}). As a result, {self.batch_size} "
-                f"eval environments will be instantiated, but only {self.n_episodes} will be used. "
-                "This might significantly slow down evaluation. To fix this, you should update your command "
-                f"to increase the number of episodes to match the batch size (e.g. `eval.n_episodes={self.batch_size}`), "
-                f"or lower the batch size (e.g. `eval.batch_size={self.n_episodes}`)."
-            )
+            self.batch_size = self.n_episodes
+
+    def _auto_batch_size(self) -> int:
+        """Pick batch_size based on CPU cores, capped by n_episodes."""
+        import math
+        import os
+
+        cpu_cores = os.cpu_count() or 4
+        # Each async env worker needs ~1 core; leave headroom for main process + inference.
+        by_cpu = max(1, math.floor(cpu_cores * 0.7))
+        return min(by_cpu, self.n_episodes, 64)


@dataclass
@@ -44,6 +44,13 @@ from lerobot.utils.constants import (
 )


+def _make_vec_env_cls(use_async: bool, n_envs: int):
+    """Return the right VectorEnv constructor."""
+    if use_async and n_envs > 1:
+        return gym.vector.AsyncVectorEnv
+    return gym.vector.SyncVectorEnv
+
+
@dataclass
 class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
    task: str | None = None
@@ -80,8 +87,9 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
        """Create {suite: {task_id: VectorEnv}}.

        Default: single-task env via gym.make(). Multi-task benchmarks override.
+        AsyncVectorEnv is the default for n_envs > 1; auto-downgraded to Sync for n_envs=1.
        """
-        env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+        env_cls = gym.vector.AsyncVectorEnv if (use_async_envs and n_envs > 1) else gym.vector.SyncVectorEnv

        if self.gym_id not in gym_registry:
            print(f"gym id '{self.gym_id}' not found, attempting to import '{self.package_name}'...")
@@ -101,12 +109,17 @@ class EnvConfig(draccus.ChoiceRegistry, abc.ABC):
        def _make_one():
            return gym.make(self.gym_id, disable_env_checker=self.disable_env_checker, **self.gym_kwargs)

+        extra_kwargs: dict = {}
+        if env_cls is gym.vector.AsyncVectorEnv:
+            extra_kwargs["context"] = "forkserver"
        try:
            from gymnasium.vector import AutoresetMode

-            vec = env_cls([_make_one for _ in range(n_envs)], autoreset_mode=AutoresetMode.SAME_STEP)
+            vec = env_cls(
+                [_make_one for _ in range(n_envs)], autoreset_mode=AutoresetMode.SAME_STEP, **extra_kwargs
+            )
        except ImportError:
-            vec = env_cls([_make_one for _ in range(n_envs)])
+            vec = env_cls([_make_one for _ in range(n_envs)], **extra_kwargs)
        return {self.type: {0: vec}}

    def get_env_processors(self):
@@ -394,7 +407,12 @@ class LiberoEnv(EnvConfig):

    @property
    def gym_kwargs(self) -> dict:
-        kwargs: dict[str, Any] = {"obs_type": self.obs_type, "render_mode": self.render_mode}
+        kwargs: dict[str, Any] = {
+            "obs_type": self.obs_type,
+            "render_mode": self.render_mode,
+            "observation_height": self.observation_height,
+            "observation_width": self.observation_width,
+        }
        if self.task_ids is not None:
            kwargs["task_ids"] = self.task_ids
        return kwargs
@@ -404,7 +422,7 @@ class LiberoEnv(EnvConfig):

        if self.task is None:
            raise ValueError("LiberoEnv requires a task to be specified")
-        env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
        return create_libero_envs(
            task=self.task,
            n_envs=n_envs,
@@ -473,7 +491,7 @@ class MetaworldEnv(EnvConfig):

        if self.task is None:
            raise ValueError("MetaWorld requires a task to be specified")
-        env_cls = gym.vector.AsyncVectorEnv if use_async_envs else gym.vector.SyncVectorEnv
+        env_cls = _make_vec_env_cls(use_async_envs, n_envs)
        return create_metaworld_envs(
            task=self.task,
            n_envs=n_envs,
@@ -29,6 +29,7 @@ from gymnasium import spaces
 from libero.libero import benchmark, get_libero_path
 from libero.libero.envs import OffScreenRenderEnv

+from lerobot.envs.utils import _LazyAsyncVectorEnv
 from lerobot.types import RobotObservation


@@ -150,7 +151,17 @@ class LiberoEnv(gym.Env):

        self.init_state_id = self.episode_index  # tie each sub-env to a fixed init state

-        self._env = self._make_envs_task(task_suite, self.task_id)
+        # Extract task metadata without allocating GPU resources (safe before fork).
+        task = task_suite.get_task(task_id)
+        self.task = task.name
+        self.task_description = task.language
+        self._task_bddl_file = os.path.join(
+            get_libero_path("bddl_files"), task.problem_folder, task.bddl_file
+        )
+        self._env: OffScreenRenderEnv | None = (
+            None  # deferred — created on first reset() inside the worker subprocess
+        )
+
        default_steps = 500
        self._max_episode_steps = (
            TASK_SUITE_MAX_STEPS.get(task_suite_name, default_steps)
@@ -221,29 +232,33 @@ class LiberoEnv(gym.Env):
            low=ACTION_LOW, high=ACTION_HIGH, shape=(ACTION_DIM,), dtype=np.float32
        )

+    def _ensure_env(self) -> None:
+        """Create the underlying OffScreenRenderEnv on first use.
+
+        Called inside the worker subprocess after fork(), so each worker gets
+        its own clean EGL context rather than inheriting a stale one from the
+        parent process (which causes EGL_BAD_CONTEXT crashes with AsyncVectorEnv).
+        """
+        if self._env is not None:
+            return
+        env = OffScreenRenderEnv(
+            bddl_file_name=self._task_bddl_file,
+            camera_heights=self.observation_height,
+            camera_widths=self.observation_width,
+        )
+        env.reset()
+        self._env = env
+
    def render(self):
+        self._ensure_env()
        raw_obs = self._env.env._get_observations()
        pixels = self._format_raw_obs(raw_obs)["pixels"]
        image = next(iter(pixels.values()))
        image = image[::-1, ::-1]  # flip both H and W for visualization
        return image

-    def _make_envs_task(self, task_suite: Any, task_id: int = 0):
-        task = task_suite.get_task(task_id)
-        self.task = task.name
-        self.task_description = task.language
-        task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file)
-
-        env_args = {
-            "bddl_file_name": task_bddl_file,
-            "camera_heights": self.observation_height,
-            "camera_widths": self.observation_width,
-        }
-        env = OffScreenRenderEnv(**env_args)
-        env.reset()
-        return env
-
    def _format_raw_obs(self, raw_obs: RobotObservation) -> RobotObservation:
+        assert self._env is not None, "_format_raw_obs called before _ensure_env()"
        images = {}
        for camera_name in self.camera_name:
            image = raw_obs[camera_name]
@@ -295,6 +310,7 @@ class LiberoEnv(gym.Env):
        )

    def reset(self, seed=None, **kwargs):
+        self._ensure_env()
        super().reset(seed=seed)
        self._env.seed(seed)
        raw_obs = self._env.reset()
@@ -321,6 +337,8 @@ class LiberoEnv(gym.Env):
        return observation, info

    def step(self, action: np.ndarray) -> tuple[RobotObservation, float, bool, bool, dict[str, Any]]:
+        self._ensure_env()
+        assert self._env is not None
        if action.ndim != 1:
            raise ValueError(
                f"Expected action to be 1-D (shape (action_dim,)), "
@@ -345,7 +363,8 @@ class LiberoEnv(gym.Env):
        return observation, reward, terminated, truncated, info

    def close(self):
-        self._env.close()
+        if self._env is not None:
+            self._env.close()


 def _make_env_fns(
@@ -428,6 +447,8 @@ def create_libero_envs(
    if task_ids_filter is not None:
        print(f"Restricting to task_ids={task_ids_filter}")

+    is_async = env_cls is gym.vector.AsyncVectorEnv
+
    out: dict[str, dict[int, Any]] = defaultdict(dict)
    for suite_name in suite_names:
        suite = _get_suite(suite_name)
@@ -436,6 +457,11 @@ def create_libero_envs(
        if not selected:
            raise ValueError(f"No tasks selected for suite '{suite_name}' (available: {total}).")

+        # All tasks in a suite share identical observation/action spaces.
+        # Probe once and reuse to avoid creating a temp env per task.
+        cached_obs_space: spaces.Space | None = None
+        cached_act_space: spaces.Space | None = None
+
        for tid in selected:
            fns = _make_env_fns(
                suite=suite,
@@ -449,8 +475,14 @@ def create_libero_envs(
                control_mode=control_mode,
                camera_name_mapping=camera_name_mapping,
            )
-            out[suite_name][tid] = env_cls(fns)
+            if is_async:
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
+                if cached_obs_space is None:
+                    cached_obs_space = lazy.observation_space
+                    cached_act_space = lazy.action_space
+                out[suite_name][tid] = lazy
+            else:
+                out[suite_name][tid] = env_cls(fns)
            print(f"Built vec env | suite={suite_name} | task_id={tid} | n_envs={n_envs}")

-    # return plain dicts for predictability
    return {suite: dict(task_map) for suite, task_map in out.items()}
@@ -25,6 +25,7 @@ import metaworld.policies as policies
 import numpy as np
 from gymnasium import spaces

+from lerobot.envs.utils import _LazyAsyncVectorEnv
 from lerobot.types import RobotObservation

 # ---- Load configuration data from the external JSON file ----
@@ -97,8 +98,9 @@ class MetaworldEnv(gym.Env):
        self.visualization_height = visualization_height
        self.camera_name = camera_name

-        self._env = self._make_envs_task(self.task)
-        self._max_episode_steps = self._env.max_path_length
+        self._env_name = self.task  # already stripped of "metaworld-" prefix above
+        self._env = None  # deferred — created on first reset() inside the worker subprocess
+        self._max_episode_steps = 500  # MT1 environments always have max_path_length=500
        self.task_description = TASK_DESCRIPTIONS[self.task]

        self.expert_policy = TASK_POLICY_MAPPING[self.task]()
@@ -136,6 +138,24 @@ class MetaworldEnv(gym.Env):

        self.action_space = spaces.Box(low=-1, high=1, shape=(ACTION_DIM,), dtype=np.float32)

+    def _ensure_env(self) -> None:
+        """Create the underlying MetaWorld env on first use.
+
+        Called inside the worker subprocess after fork(), so each worker gets
+        its own clean rendering context rather than inheriting a stale one from
+        the parent process (which causes crashes with AsyncVectorEnv).
+        """
+        if self._env is not None:
+            return
+        mt1 = metaworld.MT1(self._env_name, seed=42)
+        env = mt1.train_classes[self._env_name](render_mode="rgb_array", camera_name=self.camera_name)
+        env.set_task(mt1.train_tasks[0])
+        if self.camera_name == "corner2":
+            env.model.cam_pos[2] = [0.75, 0.075, 0.7]
+        env.reset()
+        env._freeze_rand_vec = False  # otherwise no randomization
+        self._env = env
+
    def render(self) -> np.ndarray:
        """
        Render the current environment frame.
@@ -143,26 +163,13 @@ class MetaworldEnv(gym.Env):
        Returns:
            np.ndarray: The rendered RGB image from the environment.
        """
+        self._ensure_env()
        image = self._env.render()
        if self.camera_name == "corner2":
            # Images from this camera are flipped — correct them
            image = np.flip(image, (0, 1))
        return image

-    def _make_envs_task(self, env_name: str):
-        mt1 = metaworld.MT1(env_name, seed=42)
-        env = mt1.train_classes[env_name](render_mode="rgb_array", camera_name=self.camera_name)
-        env.set_task(mt1.train_tasks[0])
-        if self.camera_name == "corner2":
-            env.model.cam_pos[2] = [
-                0.75,
-                0.075,
-                0.7,
-            ]  # corner2 position, similar to https://arxiv.org/pdf/2206.14244
-        env.reset()
-        env._freeze_rand_vec = False  # otherwise no randomization
-        return env
-
    def _format_raw_obs(self, raw_obs: np.ndarray) -> RobotObservation:
        image = None
        if self._env is not None:
@@ -209,6 +216,7 @@ class MetaworldEnv(gym.Env):
            observation (RobotObservation): The initial formatted observation.
            info (Dict[str, Any]): Additional info about the reset state.
        """
+        self._ensure_env()
        super().reset(seed=seed)

        raw_obs, info = self._env.reset(seed=seed)
@@ -232,6 +240,7 @@ class MetaworldEnv(gym.Env):
            truncated (bool): Whether the episode was truncated due to a time limit.
            info (Dict[str, Any]): Additional environment info.
        """
+        self._ensure_env()
        if action.ndim != 1:
            raise ValueError(
                f"Expected action to be 1-D (shape (action_dim,)), "
@@ -263,7 +272,8 @@ class MetaworldEnv(gym.Env):
        return observation, reward, terminated, truncated, info

    def close(self):
-        self._env.close()
+        if self._env is not None:
+            self._env.close()


 # ---- Main API ----------------------------------------------------------------
@@ -297,6 +307,9 @@ def create_metaworld_envs(

    print(f"Creating Meta-World envs | task_groups={task_groups} | n_envs(per task)={n_envs}")

+    is_async = env_cls is gym.vector.AsyncVectorEnv
+    cached_obs_space = None
+    cached_act_space = None
    out: dict[str, dict[int, Any]] = defaultdict(dict)

    for group in task_groups:
@@ -309,7 +322,14 @@ def create_metaworld_envs(
            # build n_envs factories
            fns = [(lambda tn=task_name: MetaworldEnv(task=tn, **gym_kwargs)) for _ in range(n_envs)]

-            out[group][tid] = env_cls(fns)
+            if is_async:
+                lazy = _LazyAsyncVectorEnv(fns, cached_obs_space, cached_act_space)
+                if cached_obs_space is None:
+                    cached_obs_space = lazy.observation_space
+                    cached_act_space = lazy.action_space
+                out[group][tid] = lazy
+            else:
+                out[group][tid] = env_cls(fns)

    # return a plain dict for consistency
    return {group: dict(task_map) for group, task_map in out.items()}
@@ -16,7 +16,7 @@
 import importlib.util
 import os
 import warnings
-from collections.abc import Mapping, Sequence
+from collections.abc import Callable, Mapping, Sequence
 from functools import singledispatch
 from typing import Any

@@ -29,7 +29,6 @@ from torch import Tensor

 from lerobot.configs.types import FeatureType, PolicyFeature
 from lerobot.envs.configs import EnvConfig
-from lerobot.types import RobotObservation
 from lerobot.utils.constants import OBS_ENV_STATE, OBS_IMAGE, OBS_IMAGES, OBS_STATE, OBS_STR
 from lerobot.utils.utils import get_channel_first_image_shape

@@ -130,59 +129,80 @@ def env_to_policy_features(env_cfg: EnvConfig) -> dict[str, PolicyFeature]:
    return policy_features


-def are_all_envs_same_type(env: gym.vector.VectorEnv) -> bool:
-    first_type = type(env.envs[0])  # Get type of first env
-    return all(type(e) is first_type for e in env.envs)  # Fast type check
+def _sub_env_has_attr(env: gym.vector.VectorEnv, attr: str) -> bool:
+    try:
+        env.get_attr(attr)
+        return True
+    except (AttributeError, Exception):
+        return False
+
+
+class _LazyAsyncVectorEnv:
+    """Defers AsyncVectorEnv creation until first use.
+
+    Creating all tasks' AsyncVectorEnvs upfront spawns N_tasks × n_envs worker
+    processes, all of which allocate EGL/GPU resources immediately. Since tasks
+    are evaluated sequentially, only one task's workers need to be alive at a
+    time. This wrapper stores the factory functions and creates the real
+    AsyncVectorEnv on first reset()/step()/call(), keeping peak process count = n_envs.
+    """
+
+    def __init__(
+        self,
+        env_fns: list[Callable],
+        observation_space=None,
+        action_space=None,
+    ):
+        self._env_fns = env_fns
+        self._env: gym.vector.AsyncVectorEnv | None = None
+        self.num_envs = len(env_fns)
+        if observation_space is not None and action_space is not None:
+            self.observation_space = observation_space
+            self.action_space = action_space
+        else:
+            tmp = env_fns[0]()
+            self.observation_space = tmp.observation_space
+            self.action_space = tmp.action_space
+            tmp.close()
+        self.single_observation_space = self.observation_space
+        self.single_action_space = self.action_space
+
+    def _ensure(self) -> None:
+        if self._env is None:
+            self._env = gym.vector.AsyncVectorEnv(self._env_fns, context="forkserver", shared_memory=True)
+
+    def reset(self, **kwargs):
+        self._ensure()
+        return self._env.reset(**kwargs)
+
+    def step(self, actions):
+        self._ensure()
+        return self._env.step(actions)
+
+    def call(self, name, *args, **kwargs):
+        self._ensure()
+        return self._env.call(name, *args, **kwargs)
+
+    def get_attr(self, name):
+        self._ensure()
+        return self._env.get_attr(name)
+
+    def close(self) -> None:
+        if self._env is not None:
+            self._env.close()
+            self._env = None


 def check_env_attributes_and_types(env: gym.vector.VectorEnv) -> None:
    with warnings.catch_warnings():
-        warnings.simplefilter("once", UserWarning)  # Apply filter only in this function
+        warnings.simplefilter("once", UserWarning)

-        if not (hasattr(env.envs[0], "task_description") and hasattr(env.envs[0], "task")):
+        if not (_sub_env_has_attr(env, "task_description") and _sub_env_has_attr(env, "task")):
            warnings.warn(
                "The environment does not have 'task_description' and 'task'. Some policies require these features.",
                UserWarning,
                stacklevel=2,
            )
-        if not are_all_envs_same_type(env):
-            warnings.warn(
-                "The environments have different types. Make sure you infer the right task from each environment. Empty task will be passed instead.",
-                UserWarning,
-                stacklevel=2,
-            )
-
-
-def add_envs_task(env: gym.vector.VectorEnv, observation: RobotObservation) -> RobotObservation:
-    """Adds task feature to the observation dict with respect to the first environment attribute."""
-    if hasattr(env.envs[0], "task_description"):
-        task_result = env.call("task_description")
-
-        if isinstance(task_result, tuple):
-            task_result = list(task_result)
-
-        if not isinstance(task_result, list):
-            raise TypeError(f"Expected task_description to return a list, got {type(task_result)}")
-        if not all(isinstance(item, str) for item in task_result):
-            raise TypeError("All items in task_description result must be strings")
-
-        observation["task"] = task_result
-    elif hasattr(env.envs[0], "task"):
-        task_result = env.call("task")
-
-        if isinstance(task_result, tuple):
-            task_result = list(task_result)
-
-        if not isinstance(task_result, list):
-            raise TypeError(f"Expected task to return a list, got {type(task_result)}")
-        if not all(isinstance(item, str) for item in task_result):
-            raise TypeError("All items in task result must be strings")
-
-        observation["task"] = task_result
-    else:  #  For envs without language instructions, e.g. aloha transfer cube and etc.
-        num_envs = observation[list(observation.keys())[0]].shape[0]
-        observation["task"] = ["" for _ in range(num_envs)]
-    return observation


 def _close_single_env(env: Any) -> None:
@@ -136,8 +136,8 @@ class TokenizerProcessorStep(ObservationProcessorStep):
        # Standardize to a list of strings for the tokenizer
        if isinstance(task, str):
            return [task]
-        elif isinstance(task, list) and all(isinstance(t, str) for t in task):
-            return task
+        elif isinstance(task, (list, tuple)) and all(isinstance(t, str) for t in task):
+            return list(task)

        return None

@@ -73,7 +73,6 @@ from lerobot.configs import parser
 from lerobot.configs.eval import EvalPipelineConfig
 from lerobot.envs.factory import make_env, make_env_pre_post_processors
 from lerobot.envs.utils import (
-    add_envs_task,
    check_env_attributes_and_types,
    close_envs,
    preprocess_observation,
@@ -166,9 +165,15 @@ def rollout(
        if return_observations:
            all_observations.append(deepcopy(observation))

-        # Infer "task" from attributes of environments.
-        # TODO: works with SyncVectorEnv but not AsyncVectorEnv
-        observation = add_envs_task(env, observation)
+        # Infer "task" from sub-environments (prefer natural language description).
+        # env.call() works with both SyncVectorEnv and AsyncVectorEnv.
+        try:
+            observation["task"] = list(env.call("task_description"))
+        except (AttributeError, NotImplementedError):
+            try:
+                observation["task"] = list(env.call("task"))
+            except (AttributeError, NotImplementedError):
+                observation["task"] = [""] * env.num_envs

        # Apply environment-specific preprocessing (e.g., LiberoProcessorStep for LIBERO)
        observation = env_preprocessor(observation)
@@ -318,8 +323,9 @@ def eval_policy(
        n_to_render_now = min(max_episodes_rendered - n_episodes_rendered, env.num_envs)
        if isinstance(env, gym.vector.SyncVectorEnv):
            ep_frames.append(np.stack([env.envs[i].render() for i in range(n_to_render_now)]))  # noqa: B023
-        elif isinstance(env, gym.vector.AsyncVectorEnv):
+        elif hasattr(env, "call"):
            # Here we must render all frames and discard any we don't need.
+            # Covers AsyncVectorEnv and _LazyAsyncVectorEnv (which wraps one).
            ep_frames.append(np.stack(env.call("render")[:n_to_render_now]))

    if max_episodes_rendered > 0:
@@ -521,7 +527,7 @@ def eval_main(cfg: EvalPipelineConfig):

    logging.info(colored("Output dir:", "yellow", attrs=["bold"]) + f" {cfg.output_dir}")

-    logging.info("Making environment.")
+    logging.info(f"Making environment (batch_size={cfg.eval.batch_size}, async={cfg.eval.use_async_envs}).")
    envs = make_env(
        cfg.env,
        n_envs=cfg.eval.batch_size,
@@ -755,23 +761,39 @@ def eval_policy_all(
    )

    if max_parallel_tasks <= 1:
-        # sequential path (single accumulator path on the main thread)
-        # NOTE: keeping a single-threaded accumulator avoids concurrent list appends or locks
-        for task_group, task_id, env in tasks:
-            tg, tid, metrics = task_runner(task_group, task_id, env)
-            _accumulate_to(tg, metrics)
-            per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+        prefetch_thread: threading.Thread | None = None
+        for i, (task_group, task_id, env) in enumerate(tasks):
+            if prefetch_thread is not None:
+                prefetch_thread.join()
+                prefetch_thread = None
+
+            try:
+                tg, tid, metrics = task_runner(task_group, task_id, env)
+                _accumulate_to(tg, metrics)
+                per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+            finally:
+                env.close()
+                # Prefetch next task's workers *after* closing current env to prevent
+                # GPU memory overlap between consecutive tasks.
+                if i + 1 < len(tasks):
+                    next_env = tasks[i + 1][2]
+                    if hasattr(next_env, "_ensure"):
+                        prefetch_thread = threading.Thread(target=next_env._ensure, daemon=True)
+                        prefetch_thread.start()
    else:
-        # threaded path: submit all tasks, consume completions on main thread and accumulate there
        with cf.ThreadPoolExecutor(max_workers=max_parallel_tasks) as executor:
            fut2meta = {}
            for task_group, task_id, env in tasks:
                fut = executor.submit(task_runner, task_group, task_id, env)
-                fut2meta[fut] = (task_group, task_id)
+                fut2meta[fut] = (task_group, task_id, env)
            for fut in cf.as_completed(fut2meta):
-                tg, tid, metrics = fut.result()
-                _accumulate_to(tg, metrics)
-                per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+                tg, tid, env = fut2meta[fut]
+                try:
+                    tg, tid, metrics = fut.result()
+                    _accumulate_to(tg, metrics)
+                    per_task_infos.append({"task_group": tg, "task_id": tid, "metrics": metrics})
+                finally:
+                    env.close()

    # compute aggregated metrics helper (robust to lists/scalars)
    def _agg_from_list(xs):