Added SmolVLA and LeRobot integration with configs, validation, and documentation.

configs/LeRobot/lerobot_datasource_base.toml

@@ -0,0 +1,9 @@
+
+# The training and testing dataset
+datasource = "LeRobot"
+
+# IID or non-IID? LeRobot handles deterministic episode partitioning internally.
+sampler = "iid"
+
+# The random seed used for deterministic train/test split and client partitioning.
+random_seed = 1

configs/LeRobot/smolvla_fedavg_two_client_smoke.toml

@@ -0,0 +1,70 @@
+[clients]
+
+# Type
+type = "simple"
+
+# The total number of clients
+total_clients = 2
+
+# The number of clients selected in each round
+per_round = 2
+
+# Should the clients compute test accuracy locally?
+do_test = false
+
+[server]
+address = "127.0.0.1"
+port = 8001
+simulate_wall_time = false
+checkpoint_path = "checkpoints/lerobot/smolvla_fedavg_two_client_smoke"
+model_path = "models/lerobot/smolvla_fedavg_two_client_smoke"
+
+[data]
+include = "lerobot_datasource_base.toml"
+
+[trainer]
+
+# The type of the trainer
+type = "lerobot"
+
+# The maximum number of training rounds
+rounds = 1
+
+# The maximum number of clients running concurrently
+max_concurrency = 1
+
+# The machine learning model
+model_type = "smolvla"
+model_name = "smolvla"
+
+# Number of epoches for local training in each communication round
+epochs = 1
+batch_size = 2
+optimizer = "AdamW"
+
+[algorithm]
+
+# Aggregation algorithm
+type = "fedavg"
+
+[parameters]
+
+[parameters.policy]
+type = "smolvla"
+path = "lerobot/smolvla_base"
+finetune_mode = "adapter"
+precision = "fp32"
+device = "cpu"
+
+[parameters.dataset]
+repo_id = "lerobot/pusht_image"
+delta_timestamps = { observation_image = [-0.2, -0.1, 0.0] }
+
+[parameters.transforms]
+image_size = [224, 224]
+normalize = true
+interpolation = "bilinear"
+
+[parameters.optimizer]
+lr = 0.0001
+weight_decay = 0.0

configs/LeRobot/smolvla_full_finetune.toml

@@ -0,0 +1,77 @@
+[clients]
+
+# Type
+type = "simple"
+
+# The total number of clients
+total_clients = 2
+
+# The number of clients selected in each round
+per_round = 2
+
+# Should the clients compute test accuracy locally?
+do_test = true
+
+[server]
+address = "127.0.0.1"
+port = 8002
+simulate_wall_time = false
+checkpoint_path = "checkpoints/lerobot/smolvla_full_finetune"
+model_path = "models/lerobot/smolvla_full_finetune"
+
+[data]
+include = "lerobot_datasource_base.toml"
+
+[trainer]
+
+# The type of the trainer
+type = "lerobot"
+
+# The maximum number of training rounds
+rounds = 10
+
+# The maximum number of clients running concurrently
+max_concurrency = 1
+
+# The machine learning model
+model_type = "smolvla"
+model_name = "smolvla"
+
+# Number of epoches for local training in each communication round
+epochs = 5
+
+# SmolVLA upstream fine-tuning guidance uses batch size 64.
+batch_size = 64
+optimizer = "AdamW"
+
+[algorithm]
+
+# Aggregation algorithm
+type = "fedavg"
+
+[parameters]
+
+[parameters.policy]
+type = "smolvla"
+path = "lerobot/smolvla_base"
+finetune_mode = "full"
+precision = "bf16"
+device = "cuda"
+
+[parameters.dataset]
+repo_id = "lerobot/pusht_image"
+split_seed = 7
+train_split = 0.9
+task_aware_split = true
+task_aware_partition = true
+delta_timestamps = { observation_image = [-0.2, -0.1, 0.0] }
+
+[parameters.transforms]
+image_size = [224, 224]
+center_crop = true
+normalize = true
+interpolation = "bilinear"
+
+[parameters.optimizer]
+lr = 0.0001
+weight_decay = 0.01

configs/LeRobot/smolvla_single_client_smoke.toml

@@ -0,0 +1,70 @@
+[clients]
+
+# Type
+type = "simple"
+
+# The total number of clients
+total_clients = 1
+
+# The number of clients selected in each round
+per_round = 1
+
+# Should the clients compute test accuracy locally?
+do_test = false
+
+[server]
+address = "127.0.0.1"
+port = 8000
+simulate_wall_time = false
+checkpoint_path = "checkpoints/lerobot/smolvla_single_client_smoke"
+model_path = "models/lerobot/smolvla_single_client_smoke"
+
+[data]
+include = "lerobot_datasource_base.toml"
+
+[trainer]
+
+# The type of the trainer
+type = "lerobot"
+
+# The maximum number of training rounds
+rounds = 1
+
+# The maximum number of clients running concurrently
+max_concurrency = 1
+
+# The machine learning model
+model_type = "smolvla"
+model_name = "smolvla"
+
+# Number of epoches for local training in each communication round
+epochs = 1
+batch_size = 2
+optimizer = "AdamW"
+
+[algorithm]
+
+# Aggregation algorithm
+type = "fedavg"
+
+[parameters]
+
+[parameters.policy]
+type = "smolvla"
+path = "lerobot/smolvla_base"
+finetune_mode = "adapter"
+precision = "fp32"
+device = "cpu"
+
+[parameters.dataset]
+repo_id = "lerobot/pusht_image"
+delta_timestamps = { observation_image = [-0.2, -0.1, 0.0] }
+
+[parameters.transforms]
+image_size = [224, 224]
+normalize = true
+interpolation = "bilinear"
+
+[parameters.optimizer]
+lr = 0.0001
+weight_decay = 0.0

docs/docs/configurations/parameters.md

@@ -18,3 +18,65 @@
 
 !!! example "loss_criterion"
     All the parameter settings that need to be passed as keyword parameters when initializing the loss criterion, such as `size_average`. The set of parameters permitted or needed depends on the loss criterion.
+
+## SmolVLA + LeRobot parameter contract
+
+`Config()` keeps nested keys under `[parameters]` as dot-accessible nodes. For
+the SmolVLA + LeRobot integration, define the following sections.
+
+| Config key | Purpose | Consumption path |
+| --- | --- | --- |
+| `data.datasource = "LeRobot"` | Selects the robotics datasource family. | `plato.datasources.registry.get()` chooses the datasource module. |
+| `trainer.type = "lerobot"` | Selects the robotics trainer backend. | `plato.trainers.registry.get()` chooses the trainer class. |
+| `trainer.model_type = "smolvla"` | Selects the model family. | `plato.models.registry.get()` resolves the model factory. |
+| `trainer.model_name = "smolvla"` | Selects the concrete model entry point. | `plato.models.registry.get()` resolves the model name. |
+| `parameters.policy.type` | Policy family identifier (`smolvla` in v1). | Consumed by `plato/models/smolvla.py` and `plato/trainers/lerobot.py` via `Config().parameters.policy`. |
+| `parameters.policy.path` | Pretrained policy source (Hub/local path). | Consumed by `plato/models/smolvla.py` via `Config().parameters.policy.path`. |
+| `parameters.policy.finetune_mode` | Full fine-tune vs adapter mode switch. | Consumed by `plato/trainers/lerobot.py` to decide trainable params. |
+| `parameters.policy.precision` | Runtime precision flag (`fp32`/`fp16`/`bf16`). | Consumed by `plato/trainers/lerobot.py` for dtype/autocast setup. |
+| `parameters.policy.device` | Runtime device flag (`cpu`/`cuda`/`mps`). | Consumed by `plato/trainers/lerobot.py` for device placement. |
+| `parameters.dataset.repo_id` | LeRobot dataset identifier. | Consumed by `plato/datasources/lerobot.py` dataset loader. |
+| `parameters.dataset.delta_timestamps` | Temporal window selection per modality key. | Consumed by `plato/datasources/lerobot.py` sampling/windowing logic. |
+| `parameters.transforms.*` | Image transform controls (`image_size`, `normalize`, interpolation/crop options). | Consumed by `plato/datasources/lerobot.py` preprocessing pipeline. |
+
+### Constructor-ready dictionaries
+
+SmolVLA/LeRobot components should convert section nodes to plain dictionaries
+when passing keyword arguments into constructors:
+
+```python
+from plato.config import Config
+
+cfg = Config()
+policy_kwargs = cfg.parameters.policy._asdict()
+dataset_kwargs = cfg.parameters.dataset._asdict()
+transform_kwargs = cfg.parameters.transforms._asdict()
+```
+
+### Example
+
+```toml
+[data]
+datasource = "LeRobot"
+
+[trainer]
+type = "lerobot"
+model_type = "smolvla"
+model_name = "smolvla"
+
+[parameters.policy]
+type = "smolvla"
+path = "lerobot/smolvla_base"
+finetune_mode = "full"
+precision = "bf16"
+device = "cuda"
+
+[parameters.dataset]
+repo_id = "lerobot/pusht_image"
+delta_timestamps = { observation_image = [-0.2, -0.1, 0.0] }
+
+[parameters.transforms]
+image_size = [224, 224]
+normalize = true
+interpolation = "bilinear"
+```