"The Cognitive Revolution" | AI Builders, Researchers, and Live Player Analysis Robotics Research Update, with Keerthana Gopalakrishnan and Ted Xiao of Google DeepMind
7 snips
Apr 22, 2024 
Keerthana Gopalakrishnan and Ted Xiao from Google DeepMind reveal cutting-edge advancements in AI robotics. They discuss breakthroughs enabling robots to learn from human demonstrations and handle novel objects. The duo dives into the debate on specialized versus generalist robots and emphasizes ethical considerations in robotic design. They shed light on the integration of internet-scale data in enhancing robotic learning and reflect on the challenges of training robots to navigate human environments safely and effectively.
AI Snips
Chapters
Transcript
Episode notes
Robotics Inflection Point
- Robotics is experiencing an inflection point due to internet-scale foundation models.
- Components now work 60-70% of the time, up from 20-30%, accelerating research.
RT2: Generalization through VLMs
- RT1 showed high in-domain performance with enough data; RT2 focuses on generalization.
- Training on VLMs allows robots to link internet concepts with robotics data.
Blurring Perception and Control
- Traditional robotics pipelines separate perception and control.
- RT2 blurs this line, treating both as a single problem like Visual Question Answering (VQA).