Research Questions I am interested

This list is going to be updated from time-to-time, I don’t want perfection to be my enemy now. Updated on 26 Dec 2024.

Pretext: Most people think that Robotics is a systems problem, and it has become more and more true due to increase in availability of open-source pre-trained models. And to be honest, I am not against them and think they solve most of the problems that were very difficult to solve. Largely because of the immense prior knowledge they bring to the table. However, I think there are many other interesting questions that are not primarily systems based and can have much larger impact on the field in general.

Theoretical Models of Scaling

Scaling is fascinating, and there has been some recent works in scaling robot data (Lin et al.,). While empirical models of scaling are great, can we develop theoretical models of scaling. Predict and observe the kind of data that helps the most and why it helps the most, whether the performance obtained is fundamentally limited by data (optimal compression of the data), or model capacity, or model architecture. And based on these predictions, can we increase the performance to the theoretical maximum?

Continual Learning and Memory

There has been tremendous amount of progress in continual learning over the past two years. However, as we continue to remove some of the shortterm technical bottlenecks, we will face challenges such as what information to retain, how should we deal with conflicting information, how and to what extent partial information can effect our learning algorithm or our pretrained model?

Conflict in Continual Learning

Consider a system that receives a new information in the form of RAG or human feedback. What would be the best-way to incorporate this during test-time? In-context learning is great, but what if the information recieved conflicts with the information it learned during the training behavior. How would the model behave, is there a way to know this conflict based on uncertainty estimates or neural behavior?