liujiting/Mirod-Sim-3Tasks-HighQuality-3xReal-15FPS-20260914

Mirod simulation subset: three tasks, approximately 3x real frames

This release contains simulation data only, selected for mixed training with the local Real_3tasks release. Real recordings are not included.

Folder	Task	Real reference frames	Simulation episodes	Simulation frames	Ratio
task1/dataset	Stack the small box on the other box (叠盒子)	13,785	171	41,358	3.0002
task2/dataset	Put the cup into the tray (杯子入盘)	12,361	192	37,081	2.9998
task3/dataset	Take the box out of the tray toward the robot (取盒子)	10,733	166	32,132	2.9938
Total		36,879	529	110,571	2.9982

Selection

The source is the Mirod four-task simulation release converted on 2026-09-14. This subset uses 529 distinct physical trajectories, with one rendered appearance per trajectory. Appearance variant indices are distributed across the selection; the four variants of one motion are not repeated to meet the frame budget.

Every selected trajectory passed the source physical task, grasp, placement, return-home, pregrasp, and measured-motion quality gates. Within each task, ranking prefers lower measured joint/gripper jitter, lower arm tracking residuals and longer sustained grasp. Spatial stratification preserves all occupied source spatial cells (27/27 for task1, 27/27 for task2, 22/22 for task3), while reducing concentration in dense cells. These are simulation acceptance criteria, not a claim of real-robot policy success.

Full valid episodes are retained, including task completion and return home. No new trajectory smoothing or robot-action rescaling was applied. Frame indices 0, 2, 4, ... from the 30 Hz source are selected consistently for both images, state and actions, producing 15 Hz without changing motion speed. The source invalid terminal sample remains excluded. Small frame-budget differences preserve complete trajectories.

LeRobot layout

Each taskN/dataset is an independent LeRobot v2.1 dataset, using the same numeric keys and video keys as the corresponding real dataset:

taskN/dataset/
  data/chunk-000/episode_000000.parquet
  videos/chunk-000/observation.images.environment/episode_000000.mp4
  videos/chunk-000/observation.images.left_wrist/episode_000000.mp4
  meta/info.json
  meta/tasks.jsonl
  meta/episodes.jsonl
  meta/episodes_stats.jsonl
  selection.json
  COMPLETE.json

• observation.state: float32[7], measured joint1..joint6 in radians, gripper width in meters.
• action: float32[7], absolute joint1..joint6 position targets in radians, absolute gripper target width in meters.
• timestamp: seconds at 15 Hz; frame/episode/global indices are rebuilt per task.
• Video keys: observation.images.environment, observation.images.left_wrist.
• Images retain the simulation native resolution, 224x126, in H.264 MP4 videos. The real recordings are 1280x720; use the same model resize/padding downstream.
• Task labels remain task1, task2, task3 to match the real metadata. Map these identifiers to the desired instructions consistently for both sources.

Camera geometry and mixed training

Simulation images use the renderer's rectified pinhole projection. Camera intrinsics, distortion of the original physical cameras, and the simulated target projection are provided in calibration/cameras.json. Raw real camera images should be transformed to this projection before the common model resize/padding; do not rectify this simulation subset again.

No model-specific normalization statistics are supplied. Recompute normalization for the chosen mixture and action horizon using the same action transform as training. Actions stored here are absolute targets; a delta-action transform, if used, should be applied once.

The approximately 3:1 frame count yields approximately 75% simulation and 25% real samples when the datasets are simply concatenated and frame-shuffled. Use an explicit sampler if a different mixture, such as 50:50, is intended.