5hadytru/so101_bench_sim_4_WM

SO101 Bench Sim 4 — Working Memory (WM) Sim-Real Correspondence Variant

so101_bench_sim_4_WM is the LeRobot v3.0 working memory (WM) variant of so101_bench_sim_4, prepared for the SO-101 Bench sim-real correspondence study. It uses the same simulated counterpart of the SO-101 Bench real-world “seen” object set as the base dataset, and contains exactly the same episodes, language tasks, actions, robot states, front videos, and time-varying overhead videos as sim_4.

The WM addition is observation.images.overhead_init: at every time step, it contains the overhead RGB image from the beginning of that same episode. This provides a fixed initial scene/layout reference while observation.images.overhead continues to show the current scene.

Composition

The base dataset merges, in order:

1. so101_bench_sim_2: 398 episodes and 128,784 frames.
2. so101_bench_sim_3: 487 curated episodes and 161,514 frames.

The two sources provide 263 and 323 task descriptions, with 52 shared task strings. Their merged vocabulary has 534 distinct instructions, all used by the 885 trajectories.

Property	Value
Format	LeRobot v3.0
Robot type	so101_follower
Environment	SO101 Bench simulated tabletop bin manipulation
Split	train only (0:885)
Episodes	885
Frames	290,298
Frame rate	30 FPS
Total recorded duration	9,676.6 s (2 h 41 min 16.6 s)
Episode length	143–1,166 frames; 328.02 frames on average
Distinct language tasks	534

The dataset contains simulated data rather than real-camera or real-robot recordings. Its object-set correspondence is designed for the real-world seen condition, but physical-world performance and sim-to-real transfer require their own evaluation.

Observations and actions

Each time step contains:

• observation.state: six float32 robot position values in this order: shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos, wrist_flex.pos, wrist_roll.pos, and gripper.pos.
• action: six float32 action-position values in the same order.
• observation.images.front: 640 × 480 RGB front-camera video.
• observation.images.overhead: 640 × 480 RGB overhead video for the current time step.
• observation.images.overhead_init: 640 × 480 RGB overhead video whose image is held at the first overhead observation of the current episode.
• timestamp, frame_index, episode_index, index, and task_index.

All three camera streams are AV1, yuv420p, 30 FPS video without audio. The v3.0 dataset stores multi-episode video shards, with exact episode video spans in the metadata. overhead_init is an additional fixed per-episode context; it does not replace the current-frame overhead view.

Loading

This is a LeRobot v3.0 dataset. Use a LeRobot version that supports v3.0:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("5hadytru/so101_bench_sim_4_WM")

Intended use and limitations

Use this variant for policy experiments conditioned on an episode's initial scene layout, especially sim-real correspondence studies on the SO-101 Bench seen object set. It provides a training split only and does not include rewards, per-episode success labels, or separate validation/test splits.