CoRL2026-CSI/SO101-MA-MultiTask-700epi_10fps

SO101 MA MultiTask 700epi 10fps

This private-first dataset merges seven 100-episode SO101 10 fps datasets into a single LeRobot v3.0 multitask dataset.

Dataset Summary

• Robot: so_follower
• Format: LeRobot v3.0
• Episodes: 700
• Frames: 358,006
• FPS: 10
• Task labels: 7 dataset-level task labels, 100 episodes each
• Required numeric fields: observation.state [6], action [6]
• Required image fields: observation.images.top, observation.images.left_wrist

Sources

Task index	Task label	Source
0	ma_stack_RGBblock_on_bluedish	CoRL2026-CSI/SO101-ma_stack_RGBblock_on_bluedish_10fps
1	ma_sort_RGBblock_to_matchingplate	CoRL2026-CSI/SO101-ma_sort_RGBblock_to_matchingplate_10fps
2	ma_distribute_chocolate_pie_each_plate	CoRL2026-CSI/SO101-ma_distribute_chocolate_pie_each_plate_10fps
3	cap_open_topdrawer	CoRL2026-CSI/SO101-cap_open_topdrawer_10fps
4	cap_close_pot	CoRL2026-CSI/SO101-cap_close_pot_10fps
5	cap_pnp_100ep_baseline	CoRL2026-CSI/SO101-cap_pnp_100ep_baseline_10fps
6	cap_place_spoon_between_bread_and_cereal	CoRL2026-CSI/SO101-cap_place_spoon_between_bread_and_cereal_100epi_10fps

Schema Normalization

The merge uses a minimal required schema to avoid feature conflicts across source datasets:

• Kept: observation.state, action, observation.images.top, observation.images.left_wrist, timestamp, frame_index, episode_index, index, task_index
• Dropped: optional radian, skill, subtask, end-effector, and gripper helper fields
• Camera mapping for chocolate-pie source: observation.images.realsense_topview -> observation.images.top and observation.images.wrist -> observation.images.left_wrist

Video files are copied without re-encoding. Source video codecs are mixed h264/av1. The cap_close_pot source contains 9 extra unused frames per video stream; these are preserved from the source while episode intervals still cover all dataset frames.

Validation

Local validation loaded the merged dataset with LeRobotDataset, checked 700 episodes and 358,006 data frames, verified 100 episodes per task index, and checked timestamp/frame-index regeneration per episode. It also spot-checked 2,100 merged rows against the normalized source datasets for action, observation.state, and task_index, and verified both video streams are 10 fps with sufficient episode-interval coverage.