reference/generated/torchrl.envs.IsaacLabWrapper

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

IsaacLabWrapper

torchrl.envs.IsaacLabWrapper(*args, **kwargs)

A wrapper for IsaacLab environments.

Parameters:

• env (isaaclab.envs.ManagerBasedRLEnv or equivalent) – the environment instance to wrap. ManagerBasedEnv, ManagerBasedRLEnv, DirectRLEnv and DirectMARLEnv are all supported.

• categorical_action_encoding (bool, optional) – if True, categorical specs will be converted to the TorchRL equivalent (torchrl.data.Categorical), otherwise a one-hot encoding will be used (torchrl.data.OneHot). Defaults to False.

• allow_done_after_reset (bool, optional) – if True, it is tolerated for envs to be done just after reset() is called. Defaults to True.

• native_autoreset (bool, optional) – if True, keeps Isaac Lab’s native auto-reset observations in the collector hot path and avoids the synthetic reset call in step_and_maybe_reset(). The terminal "next" observation remains unavailable and is marked with NaN; the native reset observation is cloned into the next root observation. Defaults to False.

• from_tiled_camera (bool, optional) – if True, reads pixels from an Isaac Lab tiled camera sensor and writes them under pixels_key. This is the recommended headless rendering path. Defaults to False.

• tiled_camera_name (str, optional) – Name of the sensor in env.scene.sensors. Defaults to "tiled_camera".

• tiled_camera_data_type (str, optional) – Camera data type to read. Defaults to "rgb".

• pixels_key (NestedKey, optional) – TensorDict key where pixels are written. Defaults to "pixels".

• pixels_dtype (torch.dtype, optional) – dtype used for the output pixels. If torch.uint8 is requested and the camera returns floating point data, values are scaled from [0, 1] to [0, 255]. Defaults to torch.uint8.

• pixels_channels (int, optional) – Number of channels to keep from the camera output. Defaults to 3.

For other arguments, see the torchrl.envs.GymWrapper documentation.

Refer to the Isaac Lab doc for installation instructions.

Per-index reset

Isaac Lab’s underlying envs let the caller reset an arbitrary subset of sub-environments without disturbing the others. This wrapper plumbs that capability through the standard torchrl "_reset" mask: when the tensordict passed to reset() carries a partial "_reset" boolean mask (i.e. neither all True nor all False), only the masked sub-envs are reset and the others keep their state and episode_length_buf. The transform stack (RewardSum, InitTracker, recurrent primers, VecNormV2, …) fires on the reset rows only, exactly like a normal reset.

The per-index reset path is gated on native_autoreset=True: with the default native_autoreset=False it would conflict with the VecGymEnvTransform-based obs-swap path that step_and_maybe_reset() triggers on every “done” row (this would double-reset the affected envs). Set native_autoreset=True if you want partial-reset semantics.

State-based reset

For deterministic branching from a snapshot, snapshot the scene with get_state() and restore it with env.reset(td, set_state=True, scene_state=snapshot) (or the reset_to_state() convenience, which routes through the same path). Manager-based envs only; both work in conjunction with the partial "_reset" mask.

Note

The snapshot is passed as the scene_state keyword argument rather than inside the tensordict. For a stateless env the reset state is torch-native (tensordict / torch.Tensor entries declared in state_spec), so it lives naturally in the tensordict. Isaac Lab’s scene state is an opaque, non-torch-native object: carrying it in the tensordict would mean wrapping it as a NonTensor state_spec entry and threading it through the transform and step-MDP machinery, whereas a kwarg is simpler and keeps simulator state out of the data path.

Example

>>> import argparse
>>> import gymnasium as gym
>>> import isaaclab_tasks  # noqa: F401  
>>> from isaaclab.app import add_launcher_args, launch_simulation  
>>> from isaaclab_tasks.utils import resolve_task_config  
>>> from torchrl.envs.libs.isaac_lab import IsaacLabWrapper
>>> parser = argparse.ArgumentParser()
>>> add_launcher_args(parser)  
>>> args_cli, _ = parser.parse_known_args([])
>>> env_cfg, _ = resolve_task_config("Isaac-Ant", None, overrides=[])  
>>> env_cfg.scene.num_envs = 64  
>>> with launch_simulation(env_cfg, args_cli):  
...     env = IsaacLabWrapper(gym.make("Isaac-Ant", cfg=env_cfg))
...     rollout = env.rollout(10)
>>> rollout.shape  
torch.Size([64, 10])

Note

The example uses the Isaac Lab 3.x launch API, where launch_simulation starts only the runtime the task config needs and the physics backend is a Hydra override such as overrides=["physics=ovphysx"]. Isaac Lab 2.x starts the app with isaaclab.app.AppLauncher before gym.make and registers tasks with a -v0 suffix (Isaac-Ant-v0). The wrapper supports both versions.