lidavidsh/franka-pick-100ep-genesis

Franka Pick-Cube — Flat Scene (100 episodes)

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.

Part of the Robot Synthetic Data Generation Workshop.

Kitchen scene variant: lidavidsh/franka-pick-kitchen-100ep-genesis — same task with a realistic rustic kitchen GLB environment (~209 MB).

Dataset Overview

Item	Value
Robot	Franka Panda 7-DOF
Task	Pick up a cube from a random position on the table
Scene	Flat ground plane
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

Scene Comparison

	Flat Scene (this dataset)	Kitchen Scene
Dataset	this dataset	franka-pick-kitchen-100ep-genesis
Environment	Plain ground plane	Realistic rustic kitchen (GLB mesh)
Visual complexity	Low (solid colors)	High (textured surfaces, furniture)
Size	~80 MB	~209 MB
Script	01_gen_data.py	02_gen_data_custom_scene.py

Generation Details

Item	Value
Simulator	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:

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

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

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