added tutorials

configs/examples/dev_config.json

@@ -2,8 +2,9 @@
     "dataset_mixture": {
         "datasets": [
             {
-                "repo_id": "physical-intelligence/libero"
+                "repo_id": "lerobot/droid_100"
             }
+
         ],
         "weights": [
             1.0
@@ -13,36 +14,26 @@
         "vector_resample_strategy": "nearest"
     },
     "policy": {
-        "type": "tau0",
-        "pretrained_path": "lerobot/pi0",
-        "init_strategy": "full_he_init",
+        "type": "pi05",
+        "pretrained_path": "william-yue/pi05_base",
         "n_obs_steps": 1,
         "normalization_mapping": {
             "VISUAL": "IDENTITY",
-            "STATE": "MEAN_STD",
+            "STATE": "MIN_MAX",
             "ACTION": "MEAN_STD"
         },
         "chunk_size": 50,
         "n_action_steps": 50,
         "max_state_dim": 32,
         "max_action_dim": 32,
-        "action_expert_num_cams": 1,
-        "cloud_vlm_latency_mean": 0.16,
-        "cloud_vlm_latency_std": 0.05,
-        "cloud_vlm_latency_lower": 0.10,
-        "cloud_vlm_latency_upper": 0.25,
-        "action_decoder_latency_mean": 0.032,
-        "action_decoder_latency_std": 0.010,
-        "action_decoder_latency_lower": 0.020,
-        "action_decoder_latency_upper": 0.050,
-        "tokenizer_max_length": 52,
-        "response_max_tokens": 52,
         "proj_width": 1024,
         "num_steps": 10,
+        "init_strategy": "expert_only_he_init",
         "attention_implementation": "eager",
         "freeze_vision_encoder": true,
         "train_expert_only": true,
-        "train_state_proj": true,
+        "tokenizer_max_length": 256,
+        "discrete_action_max_length": 75,
         "optimizer_lr": 2.5e-05,
         "optimizer_betas": [
             0.9,
@@ -57,17 +48,16 @@
     "resume": false,
     "seed": 1000,
     "resolution": [224, 224],
-    "num_cams": 2,
-    "action_expert_num_cams": 1,
+    "num_cams": 1,
+    "action_expert_num_cams": 0,
     "max_state_dim": 32,
     "max_action_dim": 32,
     "action_chunk": 50,
-    "frozen_actions": 25,
-    "loss_weighting": {"MSE": 1, "CE": 1},
+    "loss_weighting": {"MSE": 10, "CE": 1},
     "num_workers": 4,
-    "batch_size": 2,
+    "batch_size": 1,
     "gradient_accumulation_steps": 1,
-    "dataloader_batch_size": 2,
+    "dataloader_batch_size": 1,
     "prefetch_factor": 8,
     "steps": 100,
     "log_freq": 1,
@@ -95,11 +85,10 @@
     },
     "wandb": {
         "enable": true,
-        "entity": "wyautox-autox",
-        "project": "tau0",
+        "project": "pi05",
         "run_id": null,
         "name": null,
-        "notes": "Dev config for pi0",
+        "notes": "Dev config for pi05",
         "tags": [],
         "group": null,
         "job_type": null,

docs/source/index.rst

@@ -8,6 +8,7 @@ OpenTau documentation
    overview
    installation
    concepts
+   tutorials
 
 .. toctree::
    :maxdepth: 2

docs/source/tutorials.rst

@@ -0,0 +1,14 @@
+Tutorials
+=========
+
+This section provides step-by-step guides for common tasks in OpenTau, including training, inference, evaluation, and adding datasets.
+
+.. toctree::
+   :maxdepth: 1
+   :caption: Guides:
+
+   tutorials/training
+   tutorials/inference
+   tutorials/evaluation
+   tutorials/datasets
+

docs/source/tutorials/datasets.rst

@@ -0,0 +1,77 @@
+Datasets
+========
+
+.. note::
+   Make sure you have followed the :doc:`/installation` guide before proceeding.
+
+Building a dataset mixture 
+--------------------------
+
+You can define a dataset mixture in your configuration file using the ``dataset_mixture`` key. Here is an example:
+
+.. code-block:: json
+
+    {
+        "dataset_mixture": {
+            "datasets": [
+                {
+                    "repo_id": "physical-intelligence/libero"
+                },
+                {
+                    "repo_id": "lerobot/droid_100"
+                }
+            ],
+            "weights": [
+                0.3,
+                0.7
+            ],
+            "action_freq": 30.0,
+        },
+        ...
+    }
+
+For each new dataset, you must add an entry to ``opentau/datasets/standard_data_format_mapping.py`` to map the dataset features to the Standard Data Format (see the :ref:`Standard Data Format section <concepts/standard-data-format>` in the Concepts documentation).
+Alternatively, you can provide a custom mapping in the dataset config using the ``data_features_name_mapping`` and ``loss_type_mapping`` keys.
+For example:
+
+.. code-block:: json
+
+    {
+        "dataset_mixture": {
+            "datasets": [
+                {
+                    "repo_id": "physical-intelligence/libero"
+                    "data_features_name_mapping": {
+                        "camera0": "observation.images.exterior_image_1_left",
+                        "camera1": "observation.images.exterior_image_2_left",
+                    }
+                    "loss_type_mapping": "MSE"
+                },
+                {
+                    "repo_id": "lerobot/droid_100"
+                }
+            ],
+            "weights": [
+                0.3,
+                0.7
+            ],
+            "action_freq": 30.0,
+        },
+        ...
+    }
+
+Computing max token length for dataset mixture
+----------------------------------------------
+
+Each training config should contain a dataset mixture definition. To evaluate the maximum token length for the dataset mixture, you can run the following command:
+
+.. code-block:: bash
+
+    python lerobot/scripts/compute_max_token_length.py \
+        --target_cfg=<path/to/your/training/config.json>\
+        --output_path=outputs/stats/token_count.json \
+        --num_workers=10
+
+This will output a token count for each language key in the dataset mixture, and save it to ``outputs/stats/token_count.json``.
+
+

docs/source/tutorials/evaluation.rst

@@ -0,0 +1,49 @@
+Evaluation
+==========
+
+.. note::
+   Make sure you have followed the :doc:`/installation` guide before proceeding.
+
+Evaluating a policy in Simulation
+---------------------------------
+
+OpenTau supports evaluation in asynchronous vectorized simulation environments. To evaluate a policy in simulation, you can launch the ``lerobot/scripts/eval.py`` script with ``accelerate launch``.
+Each accelerate process will only work on its fraction of the tasks, improving throughput.
+For example, to evaluate a policy on the LIBERO 10, run:
+
+.. code-block:: bash
+
+    accelerate launch --config_file <ACCELERATE_CONFIG_PATH> lerobot/scripts/eval.py --config_path=outputs/train/tau0/checkpoints/000040/train_config.json
+
+.. note::
+   You can't pass in an DeepSpeed accelerate config file to ``eval.py`` as DeepSpeed expects optimizer and dataloader during ``accelerator.prepare()``, which we do not provide during eval. It is recommended to pass in a DDP config.
+
+.. note::
+   Make sure that the ``EnvConfig`` and ``EvalConfig`` are set to the correct values for the simulation environment in your train config file.
+
+Evaluating policy in a LIBERO environment
+-----------------------------------------
+
+OpenTau currently supports the `LIBERO benchmark <https://libero-project.github.io/main.html>`_. To evaluate the policy on the LIBERO benchmark, add the following section to the training config:
+
+.. code-block:: javascript
+
+    {
+        ...,
+        "env": {
+            "type": "libero",
+            "task": "libero_spatial",
+            "task_ids": [0, 2]
+        },
+        "eval": {
+            "n_episodes": 8,
+            "batch_size": 8
+        },
+        "eval_freq": 25,
+        ...
+    }
+
+This will run the 0th task and 2nd task in ``libero_spatial``. Each task will run for 8 simulations in parallel.
+
+When launched with accelerate, each GPU process will only work on its fraction of the tasks, improving throughput.
+

docs/source/tutorials/inference.rst

@@ -0,0 +1,17 @@
+Inference
+=========
+
+.. note::
+   Make sure you have followed the :doc:`/installation` guide before proceeding.
+
+Running inference with a trained model
+--------------------------------------
+
+To run inference on a trained model, you will need the saved checkpoint folder from training that contains at least these two files: ``train_config.json`` and ``model.safetensors``.
+If you ran the :doc:`checkpointing and resuming tutorial </tutorials/training>`, you should be able to find the checkpoint config file at ``outputs/train/pi05/checkpoints/000040/train_config.json``.
+
+To run inference, run the following command:
+
+.. code-block:: bash
+
+    python lerobot/scripts/inference.py --config_path=outputs/train/pi05/checkpoints/000040/train_config.json

docs/source/tutorials/training.rst

@@ -0,0 +1,54 @@
+Training and Checkpointing
+==========================
+
+.. note::
+   Make sure you have followed the :doc:`/installation` guide before proceeding.
+
+Distributed Training Configuration
+----------------------------------
+
+For an accelerate config example, see `this config file <../../examples/accelerate_ci_config.yaml>`_ used for our CI pipelines.
+
+To train a model, run the following command:
+
+.. code-block:: bash
+
+    accelerate launch lerobot/scripts/train.py --config_path=examples/pi05_config.json
+
+This uses the default accelerate config file at `~/.cache/huggingface/accelerate/default_config.yaml` which is set by running ``accelerate config``.
+
+Optionally, to use a specific accelerate config file (instead of the default), run:
+
+.. code-block:: bash
+
+    accelerate launch --config_file=examples/accelerate_ci_config.yaml lerobot/scripts/train.py --config_path=examples/pi05_config.json
+
+
+Checkpointing and Resuming Training
+-----------------------------------
+
+Start training and saving checkpoints:
+
+.. code-block:: bash
+
+    accelerate launch lerobot/scripts/train.py --config_path=examples/pi05_config.json --output_dir=outputs/train/pi05 --steps 40 --log_freq 5 --save_freq 20
+
+A checkpoint should be saved at step 40. The checkpoint should be saved in the directory ``outputs/train/pi05/checkpoints/000040/``.
+
+The ``model.safetensors`` file is not automatically generated by DeepSpeed's checkpointing during training. To consolidate the sharded model checkpoint files generated by DeepSpeed into a single ``model.safetensors`` file, run:
+
+.. code-block:: bash
+
+    ./convert_checkpoint.sh outputs/train/pi05/checkpoints/000040/
+
+This generates a ``model.safetensors`` file that can be used for inference or resuming training.
+
+Training can be resumed by running:
+
+.. code-block:: bash
+
+    accelerate launch lerobot/scripts/train.py --config_path=outputs/train/pi05/checkpoints/000040/train_config.json --resume=true --steps=100
+
+.. note::
+   When resuming training from a checkpoint, the training step count will continue from the checkpoint's step, but the dataloader will be reset.
+