Add SmolVLA

lerobot/__init__.py

@@ -168,12 +168,7 @@
 )
 
 # lists all available policies from `lerobot/common/policies`
-available_policies = [
-    "act",
-    "diffusion",
-    "tdmpc",
-    "vqbet",
-]
+available_policies = ["act", "diffusion", "tdmpc", "vqbet"]
 
 # lists all available robots from `lerobot/common/robot_devices/robots`
 available_robots = [

lerobot/common/policies/__init__.py

@@ -15,5 +15,6 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
 from .pi0.configuration_pi0 import PI0Config as PI0Config
+from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig

lerobot/common/policies/factory.py

@@ -27,6 +27,7 @@
 from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
 from lerobot.common.policies.pretrained import PreTrainedPolicy
+from lerobot.common.policies.smolvla.configuration_smolvla import SmolVLAConfig
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
 from lerobot.configs.policies import PreTrainedConfig
@@ -59,6 +60,10 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
         from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
 
         return PI0FASTPolicy
+    elif name == "smolvla":
+        from lerobot.common.policies.smolvla.modeling_smolvla import SmolVLAPolicy
+
+        return SmolVLAPolicy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -76,6 +81,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return PI0Config(**kwargs)
     elif policy_type == "pi0fast":
         return PI0FASTConfig(**kwargs)
+    elif policy_type == "smolvla":
+        return SmolVLAConfig(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 

lerobot/common/policies/smolvla/configuration_smolvla.py

@@ -0,0 +1,154 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.common.optim.schedulers import (
+    CosineDecayWithWarmupSchedulerConfig,
+)
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+
+@PreTrainedConfig.register_subclass("smolvla")
+@dataclass
+class SmolVLAConfig(PreTrainedConfig):
+    # Input / output structure.
+    n_obs_steps: int = 1
+    chunk_size: int = 50
+    n_action_steps: int = 50
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Shorter state and action vectors will be padded
+    max_state_dim: int = 32
+    max_action_dim: int = 32
+
+    # Image preprocessing
+    resize_imgs_with_padding: tuple[int, int] = (512, 512)
+
+    # Add empty images. Used by smolvla_aloha_sim which adds the empty
+    # left and right wrist cameras in addition to the top camera.
+    empty_cameras: int = 0
+
+    # Converts the joint and gripper values from the standard Aloha space to
+    # the space used by the pi internal runtime which was used to train the base model.
+    adapt_to_pi_aloha: bool = False
+
+    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
+    # Gripper dimensions will remain in absolute values.
+    use_delta_joint_actions_aloha: bool = False
+
+    # Tokenizer
+    tokenizer_max_length: int = 48
+
+    # Decoding
+    num_steps: int = 10
+
+    # Attention utils
+    use_cache: bool = True
+
+    # Finetuning settings
+    freeze_vision_encoder: bool = True
+    train_expert_only: bool = True
+    train_state_proj: bool = True
+
+    # Training presets
+    optimizer_lr: float = 1e-4
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-10
+    optimizer_grad_clip_norm: float = 10
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
+    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+
+    add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.
+
+    attention_mode: str = "cross_attn"
+
+    prefix_length: int = -1
+
+    pad_language_to: str = "longest"  # "max_length"
+
+    num_expert_layers: int = -1  # Less or equal to 0 is the default where the action expert has the same number of layers of VLM. Otherwise the expert have less layers.
+    num_vlm_layers: int = 16  # Number of layers used in the VLM (first num_vlm_layers layers)
+    self_attn_every_n_layers: int = 2  # Interleave SA layers each self_attn_every_n_layers
+    expert_width_multiplier: float = 0.75  # The action expert hidden size (wrt to the VLM)
+
+    min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding
+    max_period: float = 4.0
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        """Input validation (not exhaustive)."""
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.use_delta_joint_actions_aloha:
+            raise NotImplementedError(
+                "`use_delta_joint_actions_aloha` is used by smolvla for aloha real models. It is not ported yet in LeRobot."
+            )
+
+    def validate_features(self) -> None:
+        for i in range(self.empty_cameras):
+            key = f"observation.images.empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, 480, 640),
+            )
+            self.input_features[key] = empty_camera
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> list:
+        return [0]
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None