Alexander Marc Spiridonov

I am a Robotics, Systems and Control MSc. student at ETH Zurich. My interests lie broadly in the field of machine intelligence, with a focus on robotics. I am currently interested in leveraging actionless videos for robust and generalizable robot learning. Previously I worked as a Research Assistant at the Robotics Systems Lab, where I trained and deployed highly dynamic low gravity locomotion policies using Deep Reinforcement Learning. I also spent time at the Secure, Reliable, and Intelligent Systems Lab, where I worked on a comprehesive benchmarking tool for foundation models with respect to the EU AI Act. Currently I am a Graduate Research Fellow at INSAIT working with Prof. Luc Van Gool and Dr. Danda Paudel on robotics foundation models.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  Twitter  /  GitHub

Research

We introduce a generalist robot manipulation policy trained on unlabeled videos of human and robot demonstrations, enabling the use of diverse non-action-labeled data. By leveraging dynamic 3D point clouds and a self-supervised 3D dynamics predictor, the policy improves open-vocabulary task performance and learns new tasks without action labels.

We present SpaceHopper, a three-legged, small-scale robot designed for future mobile exploration of asteroids and moons. We demonstrate controlled low-gravity jumping locomotion and attitude control using Deep Reinforcement Learning in simulation, and validate the hardware in a gravity offload test stand.

We evaluated SpaceHopper, a three-legged robot for low gravity mobility, in zero gravity during parabolic flights in collaboration with the European Space Agency. During the experiments, we deployed Deep Reinforcement Learning policies for highly dynamic locomotion control.

COMPL-AI is the first technical framework able to evaluate Generative AI models on the EU AI Act 🇪🇺. The project is being actively maintained with over 100 stars on github and has been widely covered in the media (TechCrunch, Reuters, etc.).

Other Projects

Developed a novel experimental design framework for Markov Decision Processes, incorporating safe state-action constraints and a reweighted objective to estimate unknowns in unsafe regions, backed by theoretical and experimental convergence analysis.

Developed deep learning methods to enhance road network extraction from satellite imagery, incorporating topological regularizers and re-weighting approaches to improve the topological correctness of road predictions.

Research featured in