fix environment seeding

add fixes for reproducibility

only try to start env if it is closed

revision

fix normalization and data type

Improve README

Improve README

Tests are passing, Eval pretrained model works, Add gif

Update gif

Update gif

Update gif

Update gif

Update README

Update README

update minor

Update README.md

Co-authored-by: Simon Alibert <75076266+aliberts@users.noreply.github.com>

Update README.md

Co-authored-by: Simon Alibert <75076266+aliberts@users.noreply.github.com>

Address suggestions

Update thumbnail + stats

Update thumbnail + stats

Update README.md

Co-authored-by: Alexander Soare <alexander.soare159@gmail.com>

Add more comments

Add test_examples.py

lerobot/common/policies/abstract.py

@@ -9,8 +9,19 @@ class AbstractPolicy(nn.Module):
 
     The forward method should generally not be overriden as it plays the role of handling multi-step policies. See its
     documentation for more information.
+
+    Note:
+        When implementing a concrete class (e.g. `AlohaDataset`, `PushtEnv`, `DiffusionPolicy`), you need to:
+            1. set the required class attributes:
+                - for classes inheriting from `AbstractDataset`: `available_datasets`
+                - for classes inheriting from `AbstractEnv`: `name`, `available_tasks`
+                - for classes inheriting from `AbstractPolicy`: `name`
+            2. update variables in `lerobot/__init__.py` (e.g. `available_envs`, `available_datasets_per_envs`, `available_policies`)
+            3. update variables in `tests/test_available.py` by importing your new class
     """
 
+    name: str | None = None  # same name should be used to instantiate the policy in factory.py
+
     def __init__(self, n_action_steps: int | None):
         """
         n_action_steps: Sets the cache size for storing action trajectories. If None, it is assumed that a single
@@ -18,6 +29,7 @@ class AbstractPolicy(nn.Module):
             adds that dimension.
         """
         super().__init__()
+        assert self.name is not None, "Subclasses of `AbstractPolicy` should set the `name` class attribute."
         self.n_action_steps = n_action_steps
         self.clear_action_queue()
 

lerobot/common/policies/act/policy.py

@@ -42,6 +42,8 @@ def kl_divergence(mu, logvar):
 
 
 class ActionChunkingTransformerPolicy(AbstractPolicy):
+    name = "act"
+
     def __init__(self, cfg, device, n_action_steps=1):
         super().__init__(n_action_steps)
         self.cfg = cfg

lerobot/common/policies/diffusion/policy.py

@@ -13,6 +13,8 @@ from lerobot.common.utils import get_safe_torch_device
 
 
 class DiffusionPolicy(AbstractPolicy):
+    name = "diffusion"
+
     def __init__(
         self,
         cfg,

lerobot/common/policies/factory.py

@@ -3,9 +3,9 @@ def make_policy(cfg):
         raise NotImplementedError("Only diffusion policy supports rollout_batch_size > 1 for the time being.")
 
     if cfg.policy.name == "tdmpc":
-        from lerobot.common.policies.tdmpc.policy import TDMPC
+        from lerobot.common.policies.tdmpc.policy import TDMPCPolicy
 
-        policy = TDMPC(cfg.policy, cfg.device)
+        policy = TDMPCPolicy(cfg.policy, cfg.device)
     elif cfg.policy.name == "diffusion":
         from lerobot.common.policies.diffusion.policy import DiffusionPolicy
 

lerobot/common/policies/tdmpc/policy.py

@@ -87,9 +87,11 @@ class TOLD(nn.Module):
         return torch.min(Q1, Q2) if return_type == "min" else (Q1 + Q2) / 2
 
 
-class TDMPC(AbstractPolicy):
+class TDMPCPolicy(AbstractPolicy):
     """Implementation of TD-MPC learning + inference."""
 
+    name = "tdmpc"
+
     def __init__(self, cfg, device):
         super().__init__(None)
         self.action_dim = cfg.action_dim