Brendan Tidd

[Jack Collins @Jack_T_Collins]

1/ Can test-time learning act as memory for a robot policy?

Most approaches to memory in robot policies work by giving the model access to more history. Recent Robot Foundation Models (RFMs) do this increasingly well, but there is a cost:

[Rob @roblee_rl]

perpetual cup stacking @sydekickbot

[Krishan Rana @krshnrana]

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.

🧵👇

[Brendan Tidd @TiddBrendan]

There is a strong real-world focus with immediate applications on our award-winning fleet of robots. Check out our project page for more details. https://t.co/iSsjU6Ie4n
https://t.co/wKI7ZhBHAk

[Brendan Tidd @TiddBrendan]

We are currently looking for PhD students interested in Foundation Models for Robotics. https://t.co/iSsjU6Ie4n
https://t.co/wKI7ZhBHAk

[Brendan Tidd @TiddBrendan]

Themes include:
- Self-supervised learning for multimodal data streams
- Learning robust navigation models
- AI-in-the-loop reinforcement learning for training dynamic behaviours
- Neural Radiance Fields (NeRFs) to generate a data-driven and physics informed simulation