RoboPapers Ep#13 Instant Policy: In-Context Imitation Learning via Graph Diffusion
Jun 12, 2025
Edward Johns, the Director of the Robot Learning Lab at Imperial College London, and PhD student Vitalis Vosylius explore groundbreaking advances in imitation learning for robotics. They discuss how one-shot learning is transforming robot training, making it possible for machines to learn new tasks from minimal demonstrations. The conversation highlights the innovative use of graph structures for state representation and task generalization. They also delve into enhancing human-robot interaction and the critical role tactile feedback plays in achieving precision.
AI Snips
Chapters
Transcript
Episode notes
Instant Robot Learning Demo
- Vitalis demonstrated tasks to a robot which instantly learned to complete new object configurations without additional training.
- This showcase was unique as the robot performed the task immediately after the demonstration from the same video.
Humans Inspire One-Shot Learning
- Humans can learn tasks from one demonstration, indicating robots should too.
- Teaching a robot instantly is more practical than hours or days of collecting data and training.
Data Scarcity Challenges Robot Learning
- Mapping high-dimensional observations directly to low-level actions is challenging for efficient, data-scarce robot learning.
- Robotics lacks the large-scale data available in language or vision to brute force solutions.