DeepMind researchers have successfully trained two new generalist models using the Open X-Embodiment dataset. The first model, RT-1-X, is a transformer model for robot control, and the second model, RT-2-X, is a vision-language-action model that can understand visual and aural input. Both models have outperformed their predecessors, showing the ability to carry out tasks for which they had not been formally trained. The team believes that open-sourcing data and supporting collaborative learning are essential for the advancement of robotics.
Key takeaways:
- Google's DeepMind team has partnered with 33 research institutes to create a massive shared database known as "Open X-Embodiment," which is designed to advance robotics. The database contains over 500 skills and 150,000 jobs taken from 22 different kinds of robots.
- DeepMind researchers have successfully trained two new generalist models using the Open X-Embodiment dataset. The first model, RT-1-X, is a transformer model for robot control, and the second model, RT-2-X, is a vision-language-action model that can understand visual and aural input.
- Robots trained with the Open X-Embodiment dataset have shown the ability to carry out activities for which they had never received formal instruction, particularly in tasks requiring strong spatial reasoning.
- Vishal Maini, a former DeepMind researcher, has founded a new venture fund called Mythos Ventures, which aims to have an impact on the development of AI technology in various fields, including space exploration, climate change, longevity, and healthcare.