Abstract: General-purpose household robots have long been an enticing yet
elusive goal in robotics. The success of LLMs in diverse language tasks has spurred their adoption in robotics, raising the question of whether
they will drive a similar breakthrough. Two major obstacles on this
pathway are the availability of high-quality training as well as simple evaluation of models. To address these questions, I investigate the use of simulations to train embodiment agnostic VLAs and test their real-world generalization. In addition, I will review related topics in manipulation of grasp planning, visual servoing, and RL to discuss how these approaches can inspire future manipulation systems that reconcile competing demands for precision, robustness, and usability.
Bio: I am a postdoc at the University of Freiburg, where I am co-supervised by Thomas Brox and Abhinav Valada. My PhD was supervised by Thomas Brox and Wolfram Burgard. Prior to my doctorate I obtained a B.Sc. and M.Sc. in Physics at the University of Heidelberg under Björn Ommer.
Over the course of my PhD and postdoc, I have worked on computer vision for robotic manipulation, exploring approaches ranging from reinforcement learning to visual servoing. Additionally, I have investigated self-supervised learning and hand pose estimation.
Most recently, I am working to build few-shot robot imitation systems that are able to learn from human demonstrations, as well as designing robotics algorithms that easily generalize to different application settings. I have also published several articles in the field of explainable AI.
I have completed a research internship at BAIR under the guidance of Evan Shelhamer and at Symbio Robotics together with Jon Long, resulting in work published at ICLR and IROS.
(http://maxargus.com/)