Upload convert_libero_data_to_lerobot.py

convert_libero_data_to_lerobot.py

@@ -0,0 +1,96 @@
+"""
+git clone https://github.com/huggingface/lerobot.git
+cd lerobot
+pip install -e .
+
+pip install tensorflow tensorflow_datasets
+
+git clone https://huggingface.co/datasets/openvla/modified_libero_rlds
+cd modified_libero_rlds
+git lfs pull
+"""
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+import tensorflow as tf
+import tensorflow_datasets as tfds
+
+data_dir = 'modified_libero_rlds'
+REPO_NAME = 'droid-ai/libero'
+RAW_DATASET_NAMES = [
+    'libero_10_no_noops',
+    'libero_goal_no_noops',
+    'libero_object_no_noops',
+    'libero_spatial_no_noops',
+]
+
+dataset = LeRobotDataset.create(
+    repo_id=REPO_NAME,
+    robot_type='libero',
+    fps=30,
+    features={
+        'is_first': {
+            'dtype': 'bool',
+            "shape": (1,),
+        },
+        'is_last': {
+            'dtype': 'bool',
+            "shape": (1,),
+        },
+        'is_terminal': {
+            'dtype': 'bool',
+            "shape": (1,),
+        },
+        'image': {
+            'dtype': 'image',
+            'shape': (256, 256, 3),
+        },
+        'wrist_image': {
+            'dtype': 'image',
+            'shape': (256, 256, 3),
+        },
+        'state': {
+            'dtype': 'float32',
+            'shape': (8,),
+        },
+        'joint_state': {
+            'dtype': 'float32',
+            'shape': (7,),
+        },
+        'actions': {
+            'dtype': 'float32',
+            'shape': (7,),
+        },
+        'discount': {
+            'dtype': 'float32',
+            'shape': (1,),
+        },
+        'reward': {
+            'dtype': 'float32',
+            'shape': (1,),
+        },
+    },
+    image_writer_threads=128,
+    image_writer_processes=64,
+)
+
+for raw_dataset_name in RAW_DATASET_NAMES:
+    raw_dataset = tfds.load(raw_dataset_name, data_dir=data_dir, split='train')
+    raw_dataset = raw_dataset.prefetch(tf.data.AUTOTUNE)
+    for episode in raw_dataset:
+        for step in episode['steps'].as_numpy_iterator():
+            dataset.add_frame({
+                'task': step['language_instruction'].decode(),
+                'is_first': step['is_first'][None],
+                'is_last': step['is_last'][None],
+                'is_terminal': step['is_terminal'][None],
+                'image': step['observation']['image'],
+                'wrist_image': step['observation']['wrist_image'],
+                'state': step['observation']['state'],
+                'joint_state': step['observation']['joint_state'],
+                'actions': step['action'],
+                'discount': step['discount'][None],
+                'reward': step['reward'][None],
+            })
+        dataset.save_episode()
+
+dataset.push_to_hub(upload_large_folder=True)