Introducing Instant Robot Learning
Date: 2nd November 2026
Blog Post
Key Points
- Aims:
- Build robotic foundation model which learns tasks from one demonstration
- Train this model entirely in simulation
- Address issue that post-training VLAs to a new task is expensive and doesn’t generalise well (don’t respond well to new edge cases).
- Solution:
- Vision Demonstration Action model (VDA)
- VDAs are conditioned on demonstrations rather than text
- Demonstrations provide a far richer representation of a task than text
- Model has to learn-to-learn quickly rather than memorise tasks
- Model has to learn abstract representation of task… will learn the similarities between many tasks
- Benefits of Instant Learning
- Teach robot on job without expensive post-training
- sim2real is built in: can observe robot parameters in the demonstration.