Awomo-SimDataEngine: Advancing Robot Training Data Generation
A new system enhances robot training through improved scene and asset generation.
The Awomo-PhysicalRSI team introduces Awomo-SimDataEngine, a system designed to generate effective robot-training data. This system ensures that generated scenes are not only visually plausible but also support interaction and maintain physical validity.
Key Features
- Asset services provide both rigid and articulated objects.
- Scene generation includes two methods: Unravel for editable scenes from images and SimForge for environments from text.
- A graph-native harness manages construction and validation, ensuring efficient error handling.
Performance Evaluation
Evaluations demonstrate significant improvements in training outcomes, with co-training on MuJoCo-based LIBERO-Plus increasing the success rate of a World-Action Model from 77.17% to 89.43%.