RoboMME-10x

RoboMME-10x expands the RoboMME training set from 100 to 1,000 newly generated expert trajectories for each of its 16 task families. It contains 16,000 episodes and 7,636,924 frames in LeRobot v2.1 format.

Data expansion

The dataset contains newly generated simulator rollouts rather than copies of the original trajectories. For every task, we preserve the original difficulty ratio:

Difficulty	Original	RoboMME-10x
Easy	50	500
Medium	25	250
Hard	25	250
Total per task	100	1,000

Each candidate is generated with the RoboMME scene generator and expert planner. Only successful, structurally valid rollouts are retained. Failed scene generation and planning attempts are skipped, so selected seeds may contain gaps.

PatternLock receives an additional path-level check: paths already present in the original training set or in previously accepted candidates are rejected. The resulting 1,000 PatternLock trajectories have unique path signatures and do not overlap with the 98 unique paths in the original 100 training episodes.

Seed policy

PatternLock uses seeds beginning at 2,000,000. The other 15 tasks use separate namespaces defined by

seed = 3,000,000
     + task_index * 100,000
     + difficulty_offset
     + candidate_priority

where difficulty_offset is 0 for easy, 10,000 for medium, and 20,000 for hard. The task indices are:

0  BinFill             5  ButtonUnmask       10 VideoPlaceButton
1  PickXtimes          6  VideoUnmaskSwap    11 VideoPlaceOrder
2  SwingXtimes         7  ButtonUnmaskSwap   12 MoveCube
3  StopCube            8  PickHighlight      13 InsertPeg
4  VideoUnmask         9  VideoRepick        14 RouteStick

Every task contains 1,000 unique final seeds, and task namespaces do not overlap. The selected seeds were also checked against the official RoboMME train, val, and test metadata used during generation; all intersections are zero.

Seed separation prevents exact seed reuse, but does not by itself guarantee that two scenes are semantically unrelated. Exact reproduction also requires the same simulator version, assets, and expert planner.

Format

• Two 256 x 256 RGB streams: image and wrist_image
• H.264 video at 10 FPS
• Robot state, actions, demo/execution markers, and subgoal text in Parquet
• 16 task families and 134 distinct language instructions

This is generated training data. Please refer to the upstream RoboMME project for benchmark definitions, licensing, and citation information.