Rahmen:

Uncertainty estimation plays an important role in enhancing the safety and effectiveness of critical applications like autonomous driving and robotics. In the realm of robotic manipulation, the policy governing how a robot interacts with its environment and balances the exploitation-exploration dilemma hinges on this concept. A robot may adopt a conservative approach when dealing with potentially hazardous situations, ensuring safety by avoiding risks. Conversely, when encountering novel or unfamiliar scenarios, a robot can turn to an explorative mode. This flexible change is facilitated by uncertainty estimates, which provide the robot with an understanding of unknown factors.

Problemstellung:

Bayesian uncertainty estimation can be categorized into two sub-areas, namely the Bayesian inference and Ensemble, which involves a full variety of approaches such as variational inference, evidential learning, randomized prior functions, Laplace approximations and sampling-based methods such as Hamiltonian MC, etc. However, most of these paradigms are not tailored for the vision-based deep learning that involves massive state-action space, where a huge research gap exists. For these reasons, our group is now focusing on the transfer of the uncertainty estimation methods into the robot bin-picking sceneries. The main idea is a robust deep learning system that can be aware of out-ofdistribution knowledge utilizing uncertainty and carry out certain calibrations on the predictions as a self-adaptive system.

Voraussetzungen:

Solid knowledge base and experience in computer vision, deep learning.

Basic knowledge in reinforcement learning, good experiences also highly expected.

Coding skills in Python and Linux.