lidavidsh/franka-pick-100ep-genesis

---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- franka
- genesis
- pick-and-place
- synthetic
- AMD
- RDNA4
- ROCm
configs:
- config_name: default
data_files: "data/**/*.parquet"
---

# Franka Pick-Cube 100 Episodes (Genesis Synthetic)

Synthetic pick-and-place dataset generated with the Genesis simulator on **AMD RDNA4 (Radeon AI PRO R9700)**. A scripted Franka Panda arm picks up a cube from randomized table positions, recorded as two-camera video observations in LeRobot v3.0 format.

## Dataset Overview

| Item | Value |
|------|-------|
| Robot | Franka Panda 7-DOF |
| Task | Pick up a cube from a random position on the table |
| Episodes | 100 |
| Frames per episode | 135 |
| Total frames | 13,500 |
| FPS | 30 |
| Cameras | 2 (up + side), 640×480, AV1 video |
| Format | LeRobot v3.0 |
| Size | ~80 MB |

## Generation Details

| Item | Value |
|------|-------|
| Simulator | [Genesis](https://github.com/Genesis-Embodied-AI/Genesis) 0.4.5 |
| GPU | AMD Radeon AI PRO R9700 (RDNA4, gfx1201) |
| ROCm | 7.2.0 |
| Rendering | EGL + Mesa radeonsi (hardware GPU rasterization) |
| Random seed | 42 |
| Cube X range | [0.4, 0.7] m |
| Cube Y range | [-0.2, 0.2] m |
| Success rate | 100/100 = 100% |
| Generation time | 629s (~6.3s/ep) |

The dataset was generated using:

```bash
python scripts/01_gen_data.py \
--n-episodes 100 \
--repo-id local/rdna4-video-100ep \
--fps 30 \
--no-bbox-detection
```

## Features

| Feature | Shape | Type |
|---------|-------|------|
| `observation.state` | (9,) | float32 — 7 joint pos + 2 gripper pos |
| `action` | (9,) | float32 — 7 joint pos + 2 gripper pos |
| `observation.images.up` | (3, 480, 640) | video (AV1) |
| `observation.images.side` | (3, 480, 640) | video (AV1) |

## Usage

### Load with LeRobot

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("lidavidsh/franka-pick-100ep-genesis")
print(f"Episodes: {dataset.meta.total_episodes}, Frames: {len(dataset)}")

sample = dataset[0]
print(sample["observation.state"].shape) # torch.Size([9])
print(sample["observation.images.up"].shape) # torch.Size([3, 480, 640])
```

### Train SmolVLA

```bash
python scripts/02_train_vla.py \
--dataset-id lidavidsh/franka-pick-100ep-genesis \
--n-steps 2000 \
--batch-size 4 \
--num-workers 4 \
--save-dir outputs/smolvla_genesis
```

### Requirements

- `lerobot==0.4.4`
- `torchcodec>=0.2.1` (for video decoding; on ROCm, build from source to avoid CUDA dependency)
- PyTorch 2.x

## Benchmark Results (RDNA4)

| Config | Training Time | Per-step | GPU Utilization |
|--------|:---:|:---:|:---:|
| Video nw=4 | 24.5 min | 0.73 s/step | 96% |

## License

Apache 2.0