khang123452/GR1-Manipulation-IsaacSim-Rendered

GR1 Manipulation — Isaac Sim Rendered Dataset

Dataset Description

Multi-modal rendered dataset of NVIDIA GR1 humanoid robot performing tabletop manipulation tasks in Isaac Sim. Each episode is rendered as 3 synchronized MP4 videos:

• RGB — ray-traced color video
• Depth — normalized depth map (grayscale, 0–10m range)
• Segmentation — semantic segmentation with per-object class coloring

This dataset is designed as input for NVIDIA Cosmos Transfer 2.5 to generate photorealistic sim-to-real augmented training data for robot learning.

Dataset Structure

{task_name}/
  episode_NNNN_rgb.mp4      # 1280x720, 20fps, RayTracedLighting
  episode_NNNN_depth.mp4     # 1280x720, 20fps, normalized depth
  episode_NNNN_seg.mp4       # 1280x720, 20fps, semantic segmentation

Tasks

#	Task Name	Episodes	Frames/Episode
1	PnPBottleToCabinetClose	458	~120–400
2	PnPCanToDrawerClose	2	~120–400
3	PnPCupToDrawerClose	2	~120–400
4	PnPMilkToMicrowaveClose	2	~120–400
5	PnPPotatoToMicrowaveClose	2	~120–400
6	PnPWineToCabinetClose	2	~120–400
7	PosttrainPnPNovelFromCuttingboardToBasketSplitA	2	~120–400
8	PosttrainPnPNovelFromCuttingboardToCardboardboxSplitA	2	~120–400
9	PosttrainPnPNovelFromCuttingboardToPanSplitA	2	~120–400
10	PosttrainPnPNovelFromCuttingboardToPotSplitA	2	~120–400
11	PosttrainPnPNovelFromCuttingboardToTieredbasketSplitA	2	~120–400
12	PosttrainPnPNovelFromPlacematToBasketSplitA	2	~120–400
13	PosttrainPnPNovelFromPlacematToBowlSplitA	1	~120–400
14	PosttrainPnPNovelFromPlacematToPlateSplitA	2	~120–400
15	PosttrainPnPNovelFromPlacematToTieredshelfSplitA	2	~120–400
16	PosttrainPnPNovelFromPlateToBowlSplitA	2	~120–400
17	PosttrainPnPNovelFromPlateToCardboardboxSplitA	1	~120–400
18	PosttrainPnPNovelFromPlateToPanSplitA	2	~120–400
19	PosttrainPnPNovelFromPlateToPlateSplitA	1	~120–400
20	PosttrainPnPNovelFromTrayToCardboardboxSplitA	2	~120–400
21	PosttrainPnPNovelFromTrayToPlateSplitA	2	~120–400
22	PosttrainPnPNovelFromTrayToPotSplitA	2	~120–400
23	PosttrainPnPNovelFromTrayToTieredbasketSplitA	1	~120–400
24	PosttrainPnPNovelFromTrayToTieredshelfSplitA	2	~120–400

Recording Settings

Setting	Value
Camera	egoview (1st-person, attached to robot0_head_pitch)
FOV	90° vertical
Resolution	1280×720
FPS	20
Renderer	Isaac Sim RayTracedLighting, SPP=32
Lighting	Dome light, intensity=1500
USD Export	robosuite exporter, geomgroup=[0,1,0,0,0,0] (visual meshes only)

Source Data

• Robot: NVIDIA GR1 humanoid (from gr00trobosuite)
• Tasks: RoboCasa GR1 tabletop pick-and-place tasks
• Simulation: MuJoCo → USD → Isaac Sim Replicator pipeline
• Original trajectories: 1000 episodes per task in HDF5 format

Intended Use

Input for Cosmos Transfer 2.5 world generation model to produce photorealistic augmented manipulation videos for robot policy training (sim-to-real transfer).

License

CC-BY-4.0