Robert Gieselmann

I am a final year PhD student at KTH Royal Institute of Technology in Stockholm under the supervision of Florian T. Pokorny. My research is supported by WASP, the Wallenberg AI, Autonomous Systems and Software Program. Previously, I was a research scientist intern at Meta AI (FAIR) under the supervision of Mustafa Mukadam. Before my PhD studies, I worked as a research assistant within medical robotics at the Technical University of Hamburg (TUHH). I received my M.Sc. in Robotics, Cognition, Intelligence from the Technical University of Munich (TUM).

Email  /  LinkedIn  /  Google Scholar  /  Github

Research

My goal is to teach robots how to solve complex reasoning tasks from data. In particular, I am interested in the interplay between classical methods in robotics, particularly sampling-based planning algorithms, and machine learning techniques from representation learning and reinforcement learning.

We tackle the problem of long-horizon goal-reaching from visual input data by sampling-based planning in a learned embedding space. Our method employs tools from contrastive representation learning and density estimation to implement a sampling-based planner that explores solution paths within the estimated area of the latent data support.

We present a large-scale dataset for meta and mutli-task learning in the setting of prediction for physical object interaction. More specifically, we focuse on deformable object pushing tasks and feature different material properties such as complicated mechanical phenonema such as elastoplastic deformations.

We robustify the learned policies of a hierarchical reinforcement learning agent using a graph-based planning framework. We find that the lower-level value functions used for graph construction provide more accurate distance estimates compared to the ones obtained through conventional flat RL.

A collection of benchmark simulation environments for testing and developing learning algorithms for deformable object manipulation. Our benchmark features different types of tasks and physical object properties such as elasticity and plasticity.

A comparison between Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) for the problem of learning generative models of the state distributions resulting from shortest path trajectories in robot configuration spaces.

A method that improves the sample efficiency of non-holonomic mobile robot path planning by introducing learned heuristics obtained from value-based reinforcemet learning. We leveraged estimated state-action values to guide the node expansion within the Hybrid A* algorithm towards relevant goal states.