Keynote: Robot Foundation Models
summary
Demo shows a robot completing "put the eggplant in the pot" using RT2X, and a cross-embodiment dataset RTX: 34 labs, 22 robots trains a single model that averages ~50% better than lab-specific methods and yields a 3x boost on hardest language-following examples. Second‑generation VLAs add a continuous action decoder; PIO5 is pre-trained on ~10,000 hours, improves from verbal coaching, generalizes after ~100 training homes, and has a working PyTorch port.
AI-written summary · about
chapters (11)
- 0:00The rise of robot foundation
- 1:45Understanding VLA models
- 2:57Cross-embodiment datasets
- 3:54Evaluating generalist models
- 5:39Advancing VLA architectures
- 9:27Introduction to Pi0
- 9:49Emergent robot behaviors
- 10:32Pi05 and long-horizon tasks
- 12:11Verbal instruction learning
- 12:42Generalization and scaling
- 14:11Future outlook and conclusions
related talks
40mKeynote: Open-source Multimodal AI in the Wild
18mRobotics: why now?
17mBeyond Automation: AI and the Evolution of Robotic Intelligence
24mKeynote: Next-generation frameworks for Large-scale AI
32mVideo Action Models
12mKeynote: What's New in AI
49mKeynote 4: Here Come The Robots - Python and Machine Learning- 12mKeynote: What's New in AI