My work explores learning with no or minimal use of neither labeled data nor handcrafted task specific reward functions. I have worked on pose estimation using synthetic data, and now I am investigating how to guide goal-conditioned reinforcement learning with agents selecting their own goals.

In robotics it can be very expensive to gather data, and any method that reduces that cost without increasing the burden on the engineer or operator can be very valuable.

Pose Estimation

In my earlier work I address the problem of pose estimation, determining location and orientation of objects, a problem that is important when for instance making a robotic arm grasp an object. 

Our solution uses autoencoders and CAD models to train a pose estimator for multiple objects simultaneously using only synthetic data. It utilizes a multi-view approach to handle local minima in SO3-space outputs. At the time of publishing, this had similar performance to other state of the art solutions while using only RGB information, synthetic data and being both more memory efficient and faster than many other approaches.

Exploration in Reinforcement Learning

My current work examines how to make reinforcement learning agents explore efficiently. 

Many methods, such as Ɛ-greedy, have a hard time exploring when rewards are sparse. I propose to use goal-conditioned hindsight learning and intrinsic selection and evaluation of goals to guide exploration, and I am currently exploring the viability of selecting those goals in the embedding space of world models and autoencoder systems like Dreamer.