途中まで，次回からexperiment以降
Abstract
- Bringing robot learning to the general level is difficult
- Flow matching architecture
Introduction
Versatility is important: robot can achieve diverse tasks in diverse environments
Use action chunking architecture
Related works
Availability on long tasks
Overview
22 robots
task names and segmentations
Paligemma vision-language model
To generage continuous action distributions, they used flow matching
Architecture is inspired by Transfusion
\pi_0 uses conditional flow matching
Requires right dataset
Data collection
Multi-phase training procedure
Contribution
- VLM-pretraining and flow matching
- Laundry folding,clearing table…
Conclusion
- Data: 10,000 hours of dexterous manipulation data
- OXE, DROID, Bridge
- Is transferring possible (future work)
Top comments (0)