Allen Z. Ren

I am interested in developing theoretical and algorithmic techniques that allow us to formally guarantee that robots will operate reliably in diverse environments. I have focused on two directions. The first is rigorous uncertainty quantification. In order to evaluate robots' performance and formally certify them, it is important to understand and then predict when and why they succeed or fail to generalize to new environments and tasks. This requires us to rigorously examine the confidence and uncertainties of the different components of the robotic system (perception, planner, and controller), and how they reflect the real outcomes. The second is continual adaptation and recertification. As the robot is deployed in the real world, it will likely encounter situations that it has not seen during training and in-house testing. Ensuring its performance and safety requires efficient and continual adaptation of the robot policy in the wild, and also recertification of the system under such distributional shift.

My current work involves connecting the two directions in a tight loop: robots shall be recertified with rigorous uncertainty quantification after they are deployed, and uncertainty quantification of their policies identifies environments and tasks where they shall better adapt to.

Please reach out if you also find this interesting!