HuggingFaceVLA/smol-libero

This dataset was created using LeRobot.

Dataset Card for Smol-LIBERO

Dataset Summary

Smol-LIBERO is a compact version of the LIBERO benchmark, built to make experimentation fast and accessible. At just 1.79 GB (compared to ~34 GB for the full LIBERO), it contains fewer trajectories and cameras while keeping the same multimodal structure.

Each sample includes:

• —Images from two fixed cameras
• —Two types of robot state (end-effector pose + gripper, and full 7-DoF joint positions)
• —Actions (7-DoF joint commands)

This setup is especially useful for comparing low-dimensional state inputs with high-dimensional visual inputs, or combining them in multimodal training.

Dataset Structure

Data Fields

• —`observation.images.image`: 256×256×3 RGB image (camera 1)
• —`observation.images.image2`: 256×256×3 RGB image (camera 2)
• —`observation.state` (8 floats): end-effector Cartesian pose + gripper [x, y, z, roll, pitch, yaw, gripper, gripper]
• —`observation.state.joint` (7 floats): full joint angles [joint_1, …, joint_7]
• —`action` (7 floats): target joint commands

Why is it smaller than LIBERO?

• —Fewer trajectories/tasks → subset of the full benchmark
• —Only two camera views → reduced visual redundancy
• —Reduced total frames → shorter episodes or lower FPS

That’s why Smol-LIBERO is 1.79 GB instead of 34 GB.

Intended Uses

• —Quick prototyping and debugging
• —Comparing joint-space vs. Cartesian state inputs
• —Training small VLA baselines before scaling to LIBERO

Limitations

• —Smaller task and visual diversity compared to LIBERO
• —Only two fixed camera views
• —May not fully represent generalization behavior on larger benchmarks

Citation

BibTeX:

[More Information Needed]