Add gymnasium

Author: talregev

examples/smoke_gymnasium_robotics.py

@@ -0,0 +1,19 @@
+from lerobot.envs.gymnasium_robotics import GymRoboticsEnv
+import numpy as np
+
+env = GymRoboticsEnv("FetchPickAndPlace-v4")
+obs, info = env.reset()
+print({k: type(v) for k, v in obs.items()})
+print({k: v.shape for k, v in obs["images"].items()})
+print("state shape:", obs["state"].shape)
+print("goal in obs:", "goal" in obs)
+print(env.action_space)
+print(env.action_space.shape[0])
+
+done = False
+while not done:
+    action = np.zeros(env.action_space.shape, dtype=np.float32)
+    obs, reward, terminated, truncated, info = env.step(action)
+    done = terminated or truncated
+print("rollout ok")
+env.close()

pyproject.toml

@@ -133,6 +133,7 @@ aloha = ["gym-aloha>=0.1.2,<0.2.0"]
 pusht = ["gym-pusht>=0.1.5,<0.2.0", "pymunk>=6.6.0,<7.0.0"] # TODO: Fix pymunk version in gym-pusht instead
 libero = ["lerobot[transformers-dep]", "libero @ git+https://github.com/huggingface/lerobot-libero.git@main#egg=libero"]
 metaworld = ["metaworld==3.0.0"]
+gymnasium-robotics = ["gymnasium-robotics>=1.4.1"]
 
 # All
 all = [
@@ -155,6 +156,7 @@ all = [
     "lerobot[phone]",
     "lerobot[libero]",
     "lerobot[metaworld]",
+    "lerobot[gymnasium-robotics]",
 ]
 
 [project.scripts]

src/lerobot/envs/configs.py

@@ -319,3 +319,19 @@ def gym_kwargs(self) -> dict:
             "obs_type": self.obs_type,
             "render_mode": self.render_mode,
         }
+
+@EnvConfig.register_subclass("gymnasium-robotics")
+@dataclass
+class GymRoboticsEnv(EnvConfig):
+    # minimal fields the factory/CLI may expect
+    type: str = "gymnasium-robotics"
+    task: str = "fetch_pick_and_place"
+    episode_length: int | None = None
+    max_state_dim: int | None = None
+
+    seed: int | None = 0
+    image_key: str = "agentview_image"
+    
+    @property
+    def gym_kwargs(self) -> dict:
+        return {}

src/lerobot/envs/factory.py

@@ -18,7 +18,7 @@
 import gymnasium as gym
 from gymnasium.envs.registration import registry as gym_registry
 
-from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv
+from lerobot.envs.configs import AlohaEnv, EnvConfig, LiberoEnv, PushtEnv, GymRoboticsEnv
 
 
 def make_env_config(env_type: str, **kwargs) -> EnvConfig:
@@ -28,6 +28,8 @@ def make_env_config(env_type: str, **kwargs) -> EnvConfig:
         return PushtEnv(**kwargs)
     elif env_type == "libero":
         return LiberoEnv(**kwargs)
+    elif env_type == "gymnasium-robotics":
+        return GymRoboticsEnv(**kwargs)
     else:
         raise ValueError(f"Policy type '{env_type}' is not available.")
 
@@ -85,6 +87,12 @@ def make_env(
             gym_kwargs=cfg.gym_kwargs,
             env_cls=env_cls,
         )
+    elif "gymnasium-robotics" in cfg.type:
+        from lerobot.envs.gymnasium_robotics import create_gymnasium_robotics_envs
+
+        if cfg.task is None:
+            raise ValueError("Gym robotics requires a task to be specified")
+        return create_gymnasium_robotics_envs(cfg)
 
     if cfg.gym_id not in gym_registry:
         print(f"gym id '{cfg.gym_id}' not found, attempting to import '{cfg.package_name}'...")

src/lerobot/envs/gymnasium_robotics.py

@@ -0,0 +1,183 @@
+import gymnasium_robotics
+import gymnasium as gym
+import numpy as np
+from typing import Dict
+from lerobot.envs.configs import GymRoboticsEnv
+
+def create_gymnasium_robotics_envs(
+    cfg: GymRoboticsEnv,
+    n_envs: int = 1,
+    use_async_envs: bool = False,
+) -> Dict[str, Dict[int, gym.vector.VectorEnv]]:
+    """
+    Build vectorized GymRoboticsEnv(s) from the GymRoboticsEnv config and return:
+        { "<env_type>": { 0: <VectorEnv> } }
+    Minimal and consistent with make_env(...) expected return type.
+    """
+    # pull minimal fields from the config (with safe defaults)
+    task = getattr(cfg, "task", "FetchPickAndPlace-v4")
+    base_seed = getattr(cfg, "seed", 0)
+    image_key = getattr(cfg, "image_key", "agentview_image")
+    episode_length = getattr(cfg, "episode_length", None)
+    max_state_dim = getattr(cfg, "max_state_dim", None)
+
+    # per-worker factory functions
+    def _mk_one(worker_idx: int):
+        def _ctor():
+            seed = None if base_seed is None else int(base_seed) + worker_idx
+            return GymRoboticsEnv(task=task, seed=seed, image_key=image_key, max_state_dim=max_state_dim, episode_length=episode_length)
+        return _ctor
+
+    fns = [_mk_one(i) for i in range(n_envs)]
+    vec_env = gym.vector.AsyncVectorEnv(fns) if use_async_envs else gym.vector.SyncVectorEnv(fns)
+
+    # key name kept simple/flat; matches your --env.type
+    return {"gymnasium-robotics": {0: vec_env}}
+
+class GymRoboticsEnv(gym.Env):
+    """Minimal adapter: wraps a Gymnasium-Robotics env and returns a LeRobot-style obs dict."""
+    metadata = {"render_modes": ["rgb_array"], "render_fps": 80}
+
+    def __init__(
+        self, 
+        task: str, 
+        seed: int | None = 0, 
+        image_key: str = "agentview_image", 
+        episode_length: int | None = None,
+        max_state_dim: int | None = None, 
+        **make_kwargs
+    ):
+        gym.register_envs(gymnasium_robotics)
+        make_kwargs = dict(make_kwargs or {})
+        make_kwargs["render_mode"] = "rgb_array"
+        self.env = gym.make(task, max_episode_steps=1000, **make_kwargs)
+
+        self._rng = np.random.default_rng(seed)
+        self._seed = seed
+        self._image_key = image_key
+        self._max_state_dim = max_state_dim
+
+        # action space: forward from underlying env
+        self.action_space = self.env.action_space
+
+        # --- infer observation space once (do a temp reset+render) ---
+        tmp_obs, _ = self.env.reset(seed=int(self._rng.integers(0, 2**31 - 1)) if seed is not None else None)
+        frame = self.env.render()
+        obs = self._to_obs(tmp_obs, frame)
+
+        # build observation_space to match o
+        def _box_like(x, low=-np.inf, high=np.inf, dtype=np.float32):
+            x = np.asarray(x)
+            return gym.spaces.Box(low=low, high=high, shape=x.shape, dtype=dtype)
+
+        img = obs["images"][self._image_key]
+        spaces = {
+            "images": gym.spaces.Dict({self._image_key: gym.spaces.Box(low=0, high=255, shape=img.shape, dtype=np.uint8)}),
+            "state": _box_like(obs["state"]),
+            # NEW — aliases for libero-style preprocessors:
+            "agent_pos": _box_like(obs["state"]),
+            "pixels": gym.spaces.Box(low=0, high=255, shape=img.shape, dtype=np.uint8),
+        }
+        if "goal" in obs:
+            spaces["goal"] = _box_like(obs["goal"])
+        if "achieved_goal" in obs:
+            spaces["achieved_goal"] = _box_like(obs["achieved_goal"])
+
+        self.observation_space = gym.spaces.Dict(spaces)
+        # leave env in a valid state; vector wrapper will call reset() again later
+
+        # passthrough spec (if present on wrapped env)
+        self.spec = getattr(self.env, "spec", None)
+
+        max_steps = episode_length
+        if max_steps is None:
+            # determine max episode steps for upstream code that reads _max_episode_steps
+            max_steps = getattr(self.env, "_max_episode_steps", None)
+            if max_steps is None and self.spec is not None:
+                max_steps = getattr(self.spec, "max_episode_steps", None)
+
+            # try unwrapping one level if wrapped
+            if max_steps is None and hasattr(self.env, "env"):
+                inner = getattr(self.env, "env")
+                max_steps = getattr(inner, "_max_episode_steps", None)
+                if max_steps is None:
+                    inner_spec = getattr(inner, "spec", None)
+                    if inner_spec is not None:
+                        max_steps = getattr(inner_spec, "max_episode_steps", None)
+
+            # final fallback
+            if max_steps is None:
+                max_steps = 1000  # sensible default; adjust if you prefer
+
+        self._max_episode_steps = int(max_steps)
+
+
+    def reset(self, seed: int | None = None, **kwargs):
+        if seed is None and self._seed is not None:
+            seed = int(self._rng.integers(0, 2**31 - 1))
+        super().reset(seed=seed)
+        tmp_obs, info = self.env.reset(seed=seed)
+        frame = self.env.render()
+        observation = self._to_obs(tmp_obs, frame)
+        return observation, info
+
+    def step(self, action):
+        if isinstance(self.action_space, gym.spaces.Box):
+            action = np.clip(np.asarray(action, dtype=np.float32),
+                             self.action_space.low, self.action_space.high)
+        tmp_obs, reward, terminated, truncated, info = self.env.step(action)
+        frame = self.env.render()
+        obs_out = self._to_obs(tmp_obs, frame)
+        return obs_out, float(reward), bool(terminated), bool(truncated), info
+
+    def close(self):
+        self.env.close()
+
+    def render(self):
+        """Return an RGB frame (HxWx3, uint8) like Gymnasium expects."""
+        frame = self.env.render()  # underlying env created with render_mode='rgb_array'
+        if frame is None:
+            raise RuntimeError("render() returned None; ensure render_mode='rgb_array' in make().")
+        return frame.astype(np.uint8, copy=False)
+
+    # ---- helpers ----
+    @staticmethod
+    def _flat(x):
+        if x is None: return np.zeros((0,), dtype=np.float32)
+        return np.asarray(x, dtype=np.float32).reshape(-1)
+
+    def _to_obs(self, obs, frame):
+        if isinstance(obs, dict):
+            state = self._flat(obs.get("observation"))
+            desired = obs.get("desired_goal")
+            achieved = obs.get("achieved_goal")
+            rgb = frame.astype(np.uint8, copy=False)
+        elif isinstance(obs, np.ndarray) and obs.ndim == 3 and obs.shape[-1] in (1, 3):
+            # Atari-style ndarray: treat as IMAGE, not state
+            # use obs as the frame if frame is None
+            rgb_src = frame if frame is not None else obs
+            rgb = rgb_src.astype(np.uint8, copy=False)
+            # no structured state in Atari pixels; provide empty state vector
+            state = np.empty((0,), dtype=np.float32)
+            desired = achieved = None
+        else:
+            # fallback: unknown non-dict obs → treat as state only
+            state = self._flat(obs)
+            if self._max_state_dim is not None and len(state) > self._max_state_dim:
+                state = state[:self._max_state_dim]
+            desired = achieved = None
+            rgb = frame.astype(np.uint8, copy=False)
+
+        rgb = frame.astype(np.uint8, copy=False)
+
+        out = {
+            # gym original keys
+            "images": {self._image_key: rgb},
+            "state": state,
+            # aliases expected by LeRobot preprocessors
+            "agent_pos": state,   # alias for state
+            "pixels": rgb,        # alias for a single RGB view
+        }
+        if desired is not None:  out["goal"] = self._flat(desired)
+        if achieved is not None: out["achieved_goal"] = self._flat(achieved)
+        return out