# LeRobot v3.0 Conversion Plan for PiPER Picking Tests Dataset ## Executive Summary Converting piper_picking_tests from HDF5+PNG format to LeRobot v3.0 (Parquet+MP4) for VLA fine-tuning. **VERIFIED Dataset Stats (from actual files):** - Episodes: 13 - Tasks: 12 (unique picking tasks) - Total frames: 5,016 - Cameras: 2 (table_cam 800×720, wrist_cam **1280×720**) - FPS: **11-12 FPS** (verified from actual timestamps, NOT 30!) - Robot: 7-DOF arm - Current size: ~480 MB (PNG images + HDF5) - Expected output: ~150-200 MB (compressed MP4) - **Dependencies:** LeRobot v0.4.3, PyAV 15.1.0, PyTorch 2.7.1 (all installed ✅) **Data Sources:** - **State/Action/Timestamps**: HDF5 files (`observation/state`, `action`, `timestamp`) - **Images**: PNG files referenced by paths in HDF5 - **FPS**: Calculated from actual timestamp data **Strategy:** LIBERO multi-task approach with separate tasks.parquet and task_index in frames. **Reference:** See `lerobotv3_format_explanation.md` for complete v3.0 format knowledge. ## Current Format (VERIFIED) ### Actual File Organization ``` piper_picking_tests/ ├── {episode_name}_{timestamp}.hdf5 # 13 files, 58-98 KB each ├── {episode_name}_{timestamp}.json # Episode metadata (optional, not used) └── {task_name}_images/ # 12 folders (NOTE: task name, NOT episode name!) ├── observation.images.table_cam/ # PNG frames (800×720), frame_000000.png format └── observation.images.wrist_cam/ # PNG frames (1280×720), frame_000000.png format ``` **IMPORTANT:** Image folder naming uses task name only (e.g., `cleaningcloth_images`), not full episode name with timestamp! ### HDF5 Structure (VERIFIED from pencil episode) ```python observation/state # [n_frames, 7] float32 - joint angles in degrees action # [n_frames, 7] float32 - commands timestamp # [n_frames] float64 - frame timestamps in seconds episode_index # [n_frames] int64 - all same value per episode observation/images/table_cam # [n_frames] object - paths to PNG files observation/images/wrist_cam # [n_frames] object - paths to PNG files ``` **Key Finding:** HDF5 stores **PATHS** to images, not the images themselves! - Example: `b'cleaningcloth_images/observation.images.table_cam/frame_000000.png'` - Images are separate PNG files at 800×720 (table) and 1280×720 (wrist) - **Image naming:** Current format uses `frame_000000.png` (underscore), but LeRobot's `encode_video_frames` expects `frame-000000.png` (dash) - **Solution:** Copy/rename images during conversion to match required format ### Verified Episode List ```python EPISODES = { 'cleaningcloth_20251104_205021': (168 frames, 14.6s), 'fillamentroll_20251104_204834': (276 frames, 23.1s), 'gamecontroller_20251104_203816': (335 frames, 25.0s), 'hexwrench_20251104_204002': (333 frames, 24.4s), 'pencil_20251104_205415': (297 frames, 23.2s), 'scissors_20251104_204120': (290 frames, 21.0s), 'scissors_hidden_20251104_205751': (358 frames, 28.6s), 'screwdriver_20251104_203022': (324 frames, 24.8s), 'smallkey_20251104_203257': (529 frames, 39.7s), 'smallpaper_20251104_203636': (429 frames, 31.2s), 'smallwoodenstick_20251104_204353': (485 frames, 34.4s), 'thinmetaldisk_20251104_204557': (764 frames, 55.5s), 'thinmetaldisk_20251104_204721': (428 frames, 30.7s), } # Total: 5,016 frames ``` ### Image Resolution (CORRECTED) - **table_cam**: 800×720 (W×H) RGB PNG - **wrist_cam**: 1280×720 (W×H) RGB PNG ← **NOT 640×480!** - File sizes: ~387 KB (table), ~619 KB (wrist) per frame - Total images per episode: 2 × n_frames PNG files ## Target v3.0 Structure ``` piper_picking_tests_v3/ ├── meta/ │ ├── info.json # Dataset configuration │ ├── stats.json # Aggregated statistics │ ├── tasks.parquet # 12 task descriptions (LIBERO style) │ └── episodes/ │ └── chunk-000/ │ └── file-000.parquet # 13 episode metadata (NO tasks field) ├── data/ │ └── chunk-000/ │ └── file-000.parquet # All 5,016 frames (WITH task_index) └── videos/ ├── observation.images.table_cam/ │ └── chunk-000/ │ ├── file-000.mp4 # Episode 0 (cleaningcloth) │ ├── file-001.mp4 # Episode 1 (fillamentroll) │ └── ... # 13 videos total └── observation.images.wrist_cam/ └── chunk-000/ └── ... # 13 videos total ``` ### Why LIBERO Multi-Task Approach? **Chosen because:** - ✅ 12 distinct tasks (multi-task dataset) - ✅ Clean task management via tasks.parquet - ✅ Explicit task conditioning with task_index - ✅ Scalable for adding more tasks - ✅ One video per episode (flexible loading) ### Video Encoding Strategy **Using LeRobot's built-in `encode_video_frames` function (recommended):** ```python from lerobot.datasets.video_utils import encode_video_frames import shutil from pathlib import Path def prepare_and_encode_video(image_paths, output_path, fps=12, temp_dir=None): """ Prepare images and encode to MP4 using LeRobot's encode_video_frames. NOTE: encode_video_frames expects images named 'frame-XXXXXX.png' (dash, not underscore) """ temp = Path(temp_dir) if temp_dir else Path(output_path).parent / "temp_frames" temp.mkdir(parents=True, exist_ok=True) # Copy images with correct naming (frame-XXXXXX.png) for i, src_path in enumerate(image_paths): dst = temp / f"frame-{i:06d}.png" shutil.copy(src_path, dst) # Encode using LeRobot's function encode_video_frames( imgs_dir=temp, video_path=output_path, fps=fps, vcodec="libsvtav1", # AV1 codec (default) pix_fmt="yuv420p", crf=30, # Quality (lower = better, 0-51) overwrite=True ) # Cleanup temp directory shutil.rmtree(temp) ``` **✅ TESTED and VERIFIED:** - 10 frames (800×720) encoded to 0.10 MB MP4 using libsvtav1 - Video properties: 800×720, AV1 codec (libdav1d decoder) - Encoding parameters: YUV420, CRF 30, GOP 2 **Expected compression:** - PNG: ~480 MB total (all episodes, both cameras) - MP4 (libsvtav1): ~150-200 MB total (60-70% compression) - Per episode: ~6-12 MB per camera, both cameras) - MP4 (av1): ~150-200 MB total (60-70% compression) - Per episode: ~6-12 MB per camera ## Conversion Requirements ### 1. Task Language Descriptions (CRITICAL for VLA) **Current:** Task names only (`"screwdriver"`, `"scissors"`) **Required:** Natural language instructions for VLA models (SmolVLA, Pi0, XVLA) ```python TASK_LANGUAGE_MAP = { 0: "Pick up the cleaning cloth from the table.", 1: "Grasp and pick up the filament roll.", 2: "Pick up the game controller from the table.", 3: "Pick up the hex wrench tool.", 4: "Grasp and pick up the pencil.", 5: "Pick up the scissors from the table.", 6: "Find and pick up the scissors that are partially hidden.", 7: "Pick up the screwdriver from the table.", 8: "Grasp and pick up the small key.", 9: "Pick up the small piece of paper.", 10: "Pick up the small wooden stick.", 11: "Pick up the thin metal disk.", } ``` **Episode-to-task mapping:** ```python EPISODE_TO_TASK = { 'cleaningcloth_20251104_205021': 0, 'fillamentroll_20251104_204834': 1, 'gamecontroller_20251104_203816': 2, 'hexwrench_20251104_204002': 3, 'pencil_20251104_205415': 4, 'scissors_20251104_204120': 5, 'scissors_hidden_20251104_205751': 6, 'screwdriver_20251104_203022': 7, 'smallkey_20251104_203257': 8, 'smallpaper_20251104_203636': 9, 'smallwoodenstick_20251104_204353': 10, 'thinmetaldisk_20251104_204557': 11, 'thinmetaldisk_20251104_204721': 11, # Same task, different demo } ``` ### 2. Create tasks.parquet Task descriptions as DataFrame INDEX (LIBERO style): ```python import pandas as pd # Create tasks DataFrame tasks_data = {'task_index': list(range(12))} task_descriptions = list(TASK_LANGUAGE_MAP.values()) tasks_df = pd.DataFrame(tasks_data, index=task_descriptions) # Save to parquet tasks_df.to_parquet('meta/tasks.parquet') ``` ### 3. Data Parquet Schema Frame-level data with task_index: ```python { 'observation.state': float32[7], # Joint angles 'action': float32[7], # Commands 'timestamp': float32, # Frame time 'frame_index': int64, # Frame in episode 'episode_index': int64, # Which episode 'index': int64, # Global frame index 'task_index': int64, # Maps to tasks.parquet ← CRITICAL! 'next.done': bool # Last frame marker } ``` ### 4. Episode Metadata Schema NO tasks field (LIBERO approach): ```python { 'episode_index': int64, 'length': int64, # Number of frames # Data file mappings 'data/chunk_index': 0, 'data/file_index': 0, 'dataset_from_index': int64, 'dataset_to_index': int64, # Video file mappings (per camera) 'videos/observation.images.table_cam/chunk_index': 0, 'videos/observation.images.table_cam/file_index': int64, 'videos/observation.images.table_cam/from_timestamp': float, 'videos/observation.images.table_cam/to_timestamp': float, # Per-episode statistics 'stats/action/min': float32[7], 'stats/action/max': float32[7], // ... other stats } ``` ### Step 1: Setup ```python import h5py import json import pandas as pd import numpy as np from pathlib import Path from PIL import Image import shutil import tempfile # CORRECT import paths for LeRobot v0.4.3 from lerobot.datasets.lerobot_dataset import LeRobotDataset from lerobot.datasets.video_utils import encode_video_frames # Dependencies (ALL INSTALLED ✅) # - lerobot v0.4.3 (installed from local repo) # - av 15.1.0 (PyAV for video encoding) "observation.images.table_cam": { "dtype": "video", "shape": [720, 800, 3], # Height × Width × Channels "names": ["height", "width", "channel"], "video_info": { "video.fps": 12.0, # ACTUAL FPS from timestamps (not 30!) "video.codec": "libsvtav1", # CORRECT codec name "video.pix_fmt": "yuv420p", "video.is_depth_map": False, "has_audio": False } }, "observation.images.wrist_cam": { "dtype": "video", "shape": [720, 1280, 3], # Height × Width × Channels ← CORRECTED! "names": ["height", "width", "channel"], "video_info": { "video.fps": 12.0, # ACTUAL FPS from timestamps (not 30!) "video.codec": "libsvtav1", # CORRECT codec name "video.pix_fmt": "yuv420p", "video.is_depth_map": False, "has_audio": False } }, "video_info": { "video.fps": 30.0, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.is_depth_map": False, "has_audio": False } }, "observation.images.wrist_cam": { "dtype": "video", "shape": [720, 1280, 3], # Height × Width × Channels ← CORRECTED! "names": ["height", "width", "channel"], "video_info": { "video.fps": 30.0, "video.codec": "av1", "video.pix_fmt": "yuv420p", "video.is_depth_map": False, "has_audio": False } }, "observation.state": { "dtype": "float32", "shape": [7], "names": {"motors": ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]}, "fps": 30.0 }, "action": { "dtype": "float32", "shape": [7], "names": {"motors": ["joint_1", "joint_2", "joint_3", "joint_4", "joint_5", "joint_6", "joint_7"]}, "fps": 30.0 }, "episode_index": {"dtype": "int64", "shape": [1], "names": None, "fps": 30.0}, "frame_index": {"dtype": "int64", "shape": [1], "names": None, "fps": 30.0}, "timestamp": {"dtype": "float32", "shape": [1], "names": None, "fps": 30.0}, "next.done": {"dtype": "bool", "shape": [1], "names": None, "fps": 30.0}, "index": {"dtype": "int64", "shape": [1], "names": None, "fps": 30.0}, "task_index": {"dtype": "int64", "shape": [1], "names": None, "fps": 30.0}, } ``` ### Step 3: Create Tasks Parquet (LIBERO Style) ```python def create_tasks_parquet(output_dir): """Create meta/tasks.parquet with task descriptions as index.""" task_descriptions = [ "Pick up the cleaning cloth from the table.", "Grasp and pick up the filament roll.", "Pick up the game controller from the table.", "Pick up the hex wrench tool.", "Grasp and pick up the pencil.", "Pick up the scissors from the table.", "Find and pick up the scissors that are partially hidden.", "Pick up the screwdriver from the table.", "Grasp and pick up the small key.", "Pick up the small piece of paper.", "Pick up the small wooden stick.", "Pick up the thin metal disk.", ] tasks_data = {'task_index': list(range(12))} tasks_df = pd.DataFrame(tasks_data, index=task_descriptions) output_path = Path(output_dir) / 'meta' / 'tasks.parquet' output_path.parent.mkdir(parents=True, exist_ok=True) tasks_df.to_parquet(output_path) print(f"Created {output_path}") ``` ### Step 4: Main Conversion Function ```python def convert_piper_to_lerobot_v3( source_path: Path, output_path: Path, repo_id: str = "your_username/piper_picking_tests" ): """Convert PiPER dataset to LeRobot v3.0 format.""" # Episode to task mapping (from verified data) EPISODE_TO_TASK = { 'cleaningcloth_20251104_205021': 0, 'fillamentroll_20251104_204834': 1, 'gamecontroller_20251104_203816': 2, 'hexwrench_20251104_204002': 3, 'pencil_20251104_205415': 4, 'scissors_20251104_204120': 5, 'scissors_hidden_20251104_205751': 6, 'screwdriver_20251104_203022': 7, 'smallkey_20251104_203257': 8, # Create dataset dataset = LeRobotDataset.create( repo_id=repo_id, fps=12, # ACTUAL FPS from timestamp analysis features=PIPER_FEATURES, root=output_path, robot_type="piper", use_videos=True, ) # Create dataset dataset = LeRobotDataset.create( repo_id=repo_id, fps=30, features=PIPER_FEATURES, root=output_path, robot_type="piper", use_videos=True, ) # Process each episode for ep_idx, ep_name in enumerate(episodes): print(f"\nProcessing episode {ep_idx}: {ep_name}") task_idx = EPISODE_TO_TASK[ep_name] # Load HDF5 data hdf5_path = source_path / f"{ep_name}.hdf5" with h5py.File(hdf5_path, 'r') as f: # Load arrays from HDF5 states = f['observation/state'][:] actions = f['action'][:] timestamps = f['timestamp'][:] n_frames = len(timestamps) # Get image paths from HDF5 table_paths = [p.decode('utf-8') for p in f['observation/images/table_cam'][:]] wrist_paths = [p.decode('utf-8') for p in f['observation/images/wrist_cam'][:]] print(f" Frames: {n_frames}, Task: {task_idx}") # Add frames for frame_idx in range(n_frames): # Load images from paths stored in HDF5 # NOTE: Paths use task name (e.g., cleaningcloth_images), not episode name table_img_path = source_path / table_paths[frame_idx] wrist_img_path = source_path / wrist_paths[frame_idx] # Verify files exist if not table_img_path.exists(): raise FileNotFoundError(f"Missing table image: {table_img_path}") if not wrist_img_path.exists(): raise FileNotFoundError(f"Missing wrist image: {wrist_img_path}") table_img = Image.open(table_img_path) wrist_img = Image.open(wrist_img_path) frame = { "observation.state": states[frame_idx], "action": actions[frame_idx], "observation.images.table_cam": np.array(table_img), "observation.images.wrist_cam": np.array(wrist_img), "timestamp": timestamps[frame_idx], "next.done": frame_idx == n_frames - 1, "task_index": task_idx, # LIBERO approach } dataset.add_frame(frame) # Save episode (NO task parameter - we use task_index in frames) dataset.save_episode() print(f" ✓ Saved {n_frames} frames") # Create tasks.parquet create_tasks_parquet(output_path) # Finalize dataset print("\nFinalizing dataset...") dataset.finalize() print("Conversion complete!") return dataset # Usage if __name__ == "__main__": source = Path("/home/charith/projects/PiPER/piper_picking_tests") output = Path("/home/charith/projects/PiPER/piper_picking_tests_v3") dataset = convert_piper_to_lerobot_v3(source, output) ``` ### Step 5: Validation ```python def validate_dataset(dataset_path): """Validate converted dataset.""" from lerobot.common.datasets.lerobot_dataset import LeRobotDataset # Load dataset dataset = LeRobotDataset(str(dataset_path)) print(f"Total episodes: {dataset.num_episodes}") print(f"Total frames: {dataset.num_frames}") print(f"Total tasks: {len(dataset.meta.tasks) if hasattr(dataset.meta, 'tasks') else 'N/A'}") # Check tasks.parquet tasks_path = dataset_path / 'meta' / 'tasks.parquet' if tasks_path.exists(): tasks_df = pd.read_parquet(tasks_path) print(f"\nTasks parquet: {len(tasks_df)} tasks") print(tasks_df.head()) # Load sample episode sample = dataset[0] print(f"\nSample frame keys: {sample.keys()}") print(f"Task index: {sample.get('task_index', 'NOT FOUND')}") # Check video playback print(f"\nVideo shapes:") for key in sample.keys(): if 'image' in key: print(f" {key}: {sample[key].shape}") return dataset ``` ## Testing Plan ### Phase 1: Single Episode Test (30 min) ```bash # Test on screwdriver episode only python convert_script.py --episode screwdriver_20251104_203022 ``` **Validate:** - [ ] HDF5 data loads correctly (observation/state, action, timestamp) - [ ] Images load and convert to video - [ ] task_index assigned correctly - [ ] Episode metadata has file mappings - [ ] Can load with LeRobotDataset ### Phase 2: Full Conversion (1-2 hours) ```bash # Convert all 13 episodes python convert_script.py --all ``` **Validate:** - [ ] All 13 episodes present - [ ] 5,016 total frames - [ ] tasks.parquet has 12 tasks - [ ] Video quality acceptable - [ ] File sizes reasonable (~210 MB total) ### Phase 3: VLA Compatibility Test ```python # Test with VLA model loading from lerobot.common.datasets.lerobot_dataset import LeRobotDataset dataset = LeRobotDataset("path/to/piper_picking_tests_v3") # Check task conditioning sample = dataset[0] assert 'task_index' in sample print(f"Task: {dataset.meta.tasks.iloc[sample['task_index']].name}") # Try loading with VLA model # from lerobot.common.policies.vla import SmolVLA # model = SmolVLA(...) # model.select_action(sample) ``` ## Expected Outcomes ### File Structure ``` piper_picking_tests_v3/ (~150-200 MB total) ├── meta/ │ ├── info.json (~10 KB) │ ├── stats.json (~2 KB) │ ├── tasks.parquet (~5 KB) │ └── episodes/ │ └── chunk-000/ │ └── file-000.parquet (~50 KB) ├── data/ │ └── chunk-000/ │ └── file-000.parquet (~400 KB) └── videos/ ├── observation.images.table_cam/ │ └── chunk-000/ │ └── file-000.mp4 to file-012.mp4 (~75 MB total) └── observation.images.wrist_cam/ └── chunk-000/ └── file-000.mp4 to file-012.mp4 (~120 MB total) ``` ### Statistics ### Dependencies Installed ✅ - ✅ **LeRobot v0.4.3** - Installed in editable mode from `/home/charith/projects/PiPER/lerobot` - ✅ **PyAV 15.1.0** - Video encoding/decoding (downgraded from 16.0.1 for compatibility) - ✅ **PyTorch 2.7.1** - Deep learning framework with CUDA 12.6 - ✅ **torchvision 0.22.1** - Image/video processing ## Next Steps 1. ✅ **Knowledge documented** in `lerobotv3_format_explanation.md` 2. ✅ **Conversion plan created** (this file) 3. ✅ **Dependencies installed** - LeRobot v0.4.3, PyAV 15.1.0, PyTorch 2.7.1 4. ✅ **Strategy validated** - End-to-end pipeline tested with 10 frames 5. ✅ **Video encoding verified** - libsvtav1 codec produces correct output 6. ⏭️ **Implement full conversion script** 7. ⏭️ **Test on single episode** (cleaningcloth or pencil) 8. ⏭️ **Debug and refine** 9. ⏭️ **Run full conversion** (all 13 episodes) 10. ⏭️ **Validate with VLA models** 11. ⏭️ **(Optional) Push to Hugging Face Hub**) - **Metadata**: <100 KB total ### Dependencies Installed - ✅ PyAV (av) - Video encoding/decoding - ✅ OpenCV (cv2) - Already available - ✅ h5py, pillow, pandas, pyarrow - Already installed ## Next Steps 1. ✅ **Knowledge documented** in `lerobotv3_format_explanation.md` 2. ✅ **Conversion plan created** (this file) 3. ⏭️ **Implement conversion script** 4. ⏭️ **Test on single episode** 5. ⏭️ **Debug and refine** 6. ⏭️ **Run full conversion** 7. ⏭️ **Validate with VLA models** 8. ⏭️ **(Optional) Push to Hugging Face Hub** ### Critical for Success - ✅ Use LIBERO approach (tasks.parquet + task_index in frames) - ✅ Natural language task descriptions (not just labels!) - ✅ Correct HDF5 paths: `observation/state` (singular) - ✅ One video per episode (13 files per camera) - ✅ task_index in every frame - ✅ NO tasks field in episode metadata - ✅ **Correct import paths:** `lerobot.datasets.*` (NOT `lerobot.common.datasets.*`) - ✅ **Actual FPS:** 11-12 FPS (calculate from timestamps, don't assume 30) - ✅ **Image folder naming:** Uses task name only, not full episode name - ✅ **Image renaming:** Copy `frame_XXXXXX.png` → `frame-XXXXXX.png` for encode_video_frames - ✅ **Codec name:** `libsvtav1` (not just `av1`)tate` (singular) - ✅ One video per episode (13 files per camera) - ✅ task_index in every frame - ✅ NO tasks field in episode metadata ### Common Mistakes to Avoid - ❌ Using `observations/state` instead of `observation/state` - ❌ Using task names instead of language descriptions - ❌ Adding tasks field to episodes (SVLA style, not needed for LIBERO) - ❌ Forgetting task_index in frames - ❌ Consolidating all videos into one file (use one per episode for multi-task) **Last Updated:** December 10, 2025 --- ## Testing Summary (December 10, 2025) ### ✅ Pipeline Validation Complete **Test Episode:** cleaningcloth_20251104_205021 (168 frames) **Results:** 1. ✅ **All imports successful** - LeRobotDataset, encode_video_frames, PyAV, h5py, PIL 2. ✅ **HDF5 data loading** - States (168,7), Actions (168,7), Timestamps (168) all loaded correctly 3. ✅ **Image path resolution** - Successfully read paths from HDF5 and loaded PNG files 4. ✅ **Video encoding** - 10 frames encoded to 0.10 MB MP4 using libsvtav1 codec 5. ✅ **Video verification** - Output is 800×720, AV1 codec (libdav1d), playable 6. ✅ **FPS calculation** - Actual FPS is 11.53 (NOT 30 as initially assumed!) **Key Findings:** - Calculated FPS from timestamps: **11.53 FPS** (episode duration 14.6s for 168 frames) - Video codec: libsvtav1 (SVT-AV1 Encoder v3.0.0) - Encoding parameters: Preset M10, CRF 30, YUV420, 800×720 - Compression: 10 frames = 0.10 MB (excellent compression ratio) - Image paths in HDF5 use task name: `cleaningcloth_images/...` (not full episode name) **Next Action:** Create full conversion script and test with complete episode ## References - **Complete format knowledge**: `lerobotv3_format_explanation.md` - [LeRobot v3.0 Documentation](https://huggingface.co/docs/lerobot/lerobot-dataset-v3) - [Porting Datasets Guide](https://huggingface.co/docs/lerobot/porting_datasets_v3) - [DROID Example](https://github.com/huggingface/lerobot/blob/main/examples/port_datasets/port_droid.py) **Last Updated:** December 2025