5hadytru/so101_bench_real_2

SO-101 Bench Real-World Demonstrations

This is the real-world SO-101 Bench demonstration dataset described in SO-101 Bench: Measuring Object Generalization and Spatial Reasoning in Vision-Language-Action Models. It contains teleoperated demonstrations recorded with an SO-101 follower arm on a tabletop, using household objects sourced from grocery stores, thrift stores, and the author's household.

The paper describes approximately 4,400 demonstrations (about 20 hours) over approximately 30 diverse household objects. This local LeRobot v3.0 snapshot contains 4,367 episodes, 2,465,943 frames, and 1,081 distinct instructions at 30 FPS.

Tasks

The demonstrations cover four language-conditioned manipulation tasks:

• Place each object in a plastic bin.
• Place one object next to another.
• Place one object between two others.
• Move one object in a specified direction.

The paper evaluates seen objects, unseen objects from seen classes, and fully unseen objects. Its 56-object inventory is broader than the approximately 30-object teleoperation collection represented here.

Recorded fields

Each step contains six-dimensional action and observation.state position vectors in the order shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos, plus timestamps and episode/task indices. It also contains two 640 × 480 RGB video streams:

• observation.images.overhead: the global tabletop view from an iPhone;
• observation.images.front: the fixed gripper-mounted view.

The videos are AV1, yuv420p, 30 FPS, and have no audio. This snapshot is in the LeRobot v3.0 format and contains a training split only.

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("5hadytru/so101_bench_real_2")