RLinf/dreamzero-franka-pnp

DreamZero Franka Pick-and-Place

This dataset contains real-world Franka Panda pick-and-place demonstrations used for DreamZero supervised fine-tuning and real-world deployment in RLinf.

Dataset Summary

• Format: LeRobot v2.1
• Robot: Franka Panda
• Task: Pick up the object and put it into another bin
• Episodes: 140
• Frames: 9,685
• FPS: 10
• Videos: 280
• Image streams: observation.images.image, observation.images.wrist_image
• Image size: 128 x 128 RGB
• State: float32, shape [19]
• Action: float32, shape [7]

Layout

The repository keeps the standard LeRobot directory layout:

meta/
data/
videos/

Important files include:

• meta/info.json
• meta/tasks.jsonl
• meta/stats.json
• meta/episodes.jsonl
• data/chunk-000/episode_*.parquet
• videos/chunk-000/observation.images.image/episode_*.mp4
• videos/chunk-000/observation.images.wrist_image/episode_*.mp4

Usage With RLinf

Download the dataset:

huggingface-cli download RLinf/dreamzero-franka-pnp --repo-type dataset --local-dir ./franka_pnp

Generate DreamZero metadata:

python toolkits/lerobot/generate_dreamzero_metadata.py \
  --preset franka_pnp \
  --dataset-root ./franka_pnp \
  --output-metadata ./franka_pnp_metadata.json

Use the dataset with embodiment_tag: franka_pnp. For real-world Franka evaluation in RLinf, use the DreamZero Franka config and keep the gripper action convention aligned with gripper_action_mode: signed.

Safety and Scope

This dataset is intended for research on robot imitation learning and real-world deployment. Any robot execution should be validated in a controlled environment with appropriate hardware safety checks.