This work will be presented at the Robotics: Science and Systems (RSS) conference in Los Angeles.

Project Page: imle-policy.github.io
Paper: roboticsproceedings.org/rss21/p158.pdf
Code and Datasets: github.com/krishanrana/...

IMLE Policy is highly multimodal, and we introduce consistency as a method to ensure consistent mode selection across tasks.

Single-step inference makes it easy to batch generate multiple proposals from which the most consistent with the previous action chunk is selected.

Particularly in the low data regime, we show that IMLE Policy can consistently outperform Diffusion Policy and Flow Matching.

We extended this approach to the conditional action generation setting and found that these attributes carried over, enabling us to learn performant visuomotor policies from as little as 17 demonstrations.

Are Diffusion and Flow Matching the best generative modelling algorithms for behaviour cloning in robotics?

✅Multimodality
❌Fast, Single-Step Inference
❌Sample Efficient

💡 We introduce IMLE Policy, a novel behaviour cloning approach that can satisfy all the above.

🧵👇