Finetuning

.gitmodules

@@ -3,4 +3,7 @@
 	url = https://github.com/Physical-Intelligence/aloha.git
 [submodule "third_party/libero"]
 	path = third_party/libero
-	url = https://github.com/Lifelong-Robot-Learning/LIBERO.git
+	url = https://github.com/szhaovas/LIBERO.git
+[submodule "third_party/SimplerEnv"]
+	path = third_party/SimplerEnv
+	url = https://github.com/simpler-env/SimplerEnv.git

convert_google_robot_to_lerobot.py

@@ -0,0 +1,74 @@
+import numpy as np
+import json
+
+from pathlib import Path
+
+from lerobot.common.datasets.lerobot_dataset import HF_LEROBOT_HOME
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+import tensorflow_datasets as tfds
+import tyro
+
+google_demo_conversion_cfg = {
+    "repo_id": "hchen/google_robot",
+    "demo_dir": "./simpler_env_demos/demo_collection/collected_data"
+}
+
+def main(data_cfg):
+    demo_root_dir = Path(data_cfg["demo_dir"])
+    task_dirs = [d for d in demo_root_dir.iterdir() if d.is_dir()]
+    task_metadata_paths = [d / "metadata.json" for d in task_dirs]
+
+    task_metadata = [json.loads(p.read_text()) for p in task_metadata_paths]
+
+    dataset = LeRobotDataset.create(
+        repo_id=data_cfg["repo_id"],
+        robot_type="google_robot",
+        fps=10,
+        features={
+            "image": {
+                "dtype": "image",
+                "shape": (224, 224, 3),
+                "names": ["height", "width", "channel"]
+            },
+            "state": {
+                "dtype": "float32",
+                "shape": (8,),
+                "names": ["state"],
+            },
+            "actions": {
+                "dtype": "float32",
+                "shape": (7,),
+                "names": ["actions"],
+            }
+        },
+        image_writer_threads=10,
+        image_writer_processes=5
+    )
+
+    for single_task_ds in task_metadata:
+        success_episode_files = ["./simpler_env_demos/demo_collection/" + file for file in single_task_ds["success_episode_files"]]
+        for i, episode_file in enumerate(success_episode_files):
+            traj_data = np.load(episode_file)
+
+            traj_images = traj_data["images"]
+            traj_states = traj_data["states"]
+            traj_actions = traj_data["actions"]
+            traj_language = str(traj_data["language"])
+
+            episode_len = len(traj_images)
+
+            for t in range(episode_len):
+                dataset.add_frame(
+                    {
+                        "image": traj_images[t],
+                        "state": traj_states[t],
+                        "actions": traj_actions[t],
+                        "task": traj_language
+                    }
+                )
+
+            dataset.save_episode()
+            print(f"Saved trajectory {i} out of {single_task_ds['total_saved_episodes']} for {single_task_ds['environment_name']}")
+
+if __name__ == "__main__":
+    main(data_cfg=google_demo_conversion_cfg)

create_lerobot_ds.py

@@ -0,0 +1,106 @@
+"""
+Minimal example script for converting a dataset to LeRobot format.
+
+We use the Libero dataset (stored in RLDS) for this example, but it can be easily
+modified for any other data you have saved in a custom format.
+
+Usage:
+uv run examples/libero/convert_libero_data_to_lerobot.py --data_dir /path/to/your/data
+
+If you want to push your dataset to the Hugging Face Hub, you can use the following command:
+uv run examples/libero/convert_libero_data_to_lerobot.py --data_dir /path/to/your/data --push_to_hub
+
+Note: to run the script, you need to install tensorflow_datasets:
+`uv pip install tensorflow tensorflow_datasets`
+
+You can download the raw Libero datasets from https://huggingface.co/datasets/openvla/modified_libero_rlds
+The resulting dataset will get saved to the $HF_LEROBOT_HOME directory.
+Running this conversion script will take approximately 30 minutes.
+"""
+
+import shutil
+
+from lerobot.common.datasets.lerobot_dataset import HF_LEROBOT_HOME
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+import tensorflow_datasets as tfds
+import tyro
+
+import numpy as np
+import pickle
+
+import os, struct
+
+def main(scheduler_pkl_path="./test_logs/scheduler_00000010.pkl",
+         repo_name="hchen/libero"): 
+    # Clean up any existing dataset in the output directory
+    output_path = HF_LEROBOT_HOME / repo_name
+    if output_path.exists():
+        shutil.rmtree(output_path)
+
+    # Create LeRobot dataset, define features to store
+    # OpenPi assumes that proprio is stored in `state` and actions in `action`
+    # LeRobot assumes that dtype of image data is `image`
+    dataset = LeRobotDataset.create(
+        repo_id=repo_name,
+        robot_type="panda",
+        fps=10,
+        features={
+            "image": {
+                "dtype": "image",
+                "shape": (256, 256, 3),
+                "names": ["height", "width", "channel"],
+            },
+            "wrist_image": {
+                "dtype": "image",
+                "shape": (256, 256, 3),
+                "names": ["height", "width", "channel"],
+            },
+            "state": {
+                "dtype": "float64",
+                "shape": (8,),
+                "names": ["state"],
+            },
+            "actions": {
+                "dtype": "float64",
+                "shape": (7,),
+                "names": ["actions"],
+            },
+        },
+        image_writer_threads=10,
+        image_writer_processes=5,
+    )
+
+    with open(scheduler_pkl_path, "rb") as f:
+        scheduler = pickle.load(f)
+
+    archive = scheduler.archive
+    all_trajectories = archive.data("trajectories")
+
+    include_failures = True
+
+    for elite_trajectories in all_trajectories:
+        for traj_id, traj in enumerate(elite_trajectories):
+            episode_len = np.array(traj["image"]).shape[0]
+
+            if not include_failures and not traj["success"]:
+                continue
+
+            for step in range(episode_len):
+                dataset.add_frame(
+                    {
+                        "image": traj["image"][step],
+                        "wrist_image": traj["wrist_image"][step],
+                        "state": traj["state"][step],
+                        "actions": traj["action"][step],
+                        "task": traj["prompt"]
+                    }
+                )
+
+            dataset.save_episode()
+            print(f"Saved trajectory {traj_id}!")
+
+        print(dataset)
+
+if __name__ == "__main__":
+    main(scheduler_pkl_path="./test_logs/scheduler_00000010.pkl",
+         repo_name="hchen/libero")

eval_base_libero_logs/summary.csv

@@ -0,0 +1,51 @@
+env_num,success_rate
+0,1.0
+1,1.0
+2,1.0
+3,1.0
+4,0.8
+5,1.0
+6,1.0
+7,1.0
+8,1.0
+9,1.0
+10,1.0
+11,1.0
+12,1.0
+13,1.0
+14,1.0
+15,1.0
+16,1.0
+17,1.0
+18,1.0
+19,1.0
+20,1.0
+21,0.8
+22,1.0
+23,1.0
+24,1.0
+25,1.0
+26,1.0
+27,1.0
+28,1.0
+29,1.0
+30,0.8
+31,1.0
+32,1.0
+33,1.0
+34,0.8
+35,1.0
+36,1.0
+37,1.0
+38,0.8
+39,1.0
+40,1.0
+41,1.0
+42,0.8
+43,1.0
+44,1.0
+45,1.0
+46,1.0
+47,0.8
+48,1.0
+49,1.0

eval_finetuned_libero_logs/summary.csv

@@ -0,0 +1,51 @@
+env_num,success_rate
+0,1.0
+1,1.0
+2,1.0
+3,1.0
+4,1.0
+5,1.0
+6,1.0
+7,0.8
+8,1.0
+9,0.8
+10,1.0
+11,1.0
+12,1.0
+13,0.8
+14,1.0
+15,1.0
+16,0.8
+17,1.0
+18,1.0
+19,0.8
+20,1.0
+21,0.8
+22,1.0
+23,0.8
+24,1.0
+25,1.0
+26,1.0
+27,1.0
+28,1.0
+29,1.0
+30,1.0
+31,1.0
+32,0.8
+33,1.0
+34,1.0
+35,1.0
+36,1.0
+37,1.0
+38,1.0
+39,1.0
+40,1.0
+41,0.8
+42,0.8
+43,1.0
+44,1.0
+45,1.0
+46,1.0
+47,1.0
+48,0.8
+49,1.0