← AI Terminology
Imitation Learning
Imitation learning trains agents from expert demonstrations (state-action data) rather than from sparse reward discovery alone.
It includes behavioural cloning and inverse RL families.
It includes behavioural cloning and inverse RL families.
Why It Matters in AI
Designing rewards is hard; copying experts is often easier. Robotics, self-driving, and game AI rely heavily on demonstration data. Foundation VLA models continue this tradition at scale.
Key Points
| Aspect | Description |
|---|---|
| Con | Distribution shift when agent deviates from expert states |
| Pro | Bypasses reward engineering |
| Use | Robot manipulation, autonomous driving stacks |
| Data | Expert trajectories of states and actions |
| Methods | Behavioral cloning, DAgger, IRL, GAIL |
| Related | Offline RL, sim-to-real |
Simple Analogy
Learning to cook by watching a chef’s hands rather than inventing cuisine from a vague “make it tasty” score.
Common Usage Examples
- Behavioral clone a policy network on demos
- DAgger iterative correction
- Robot teleop datasets
- VLA models trained on human videos
Summary
In short: Imitation learning teaches agents from expert demonstrations — often easier than crafting rewards from scratch.