Hello! I am a 4th-year PhD student at the University of Amsterdam, part of AMLAB. My research focuses on building foundation models (FM) for sequential decision-making (SDM) and exploring their applications in scientific research itself.

Foundation Models for Sequential Decision Making

I believe the path to creating effective foundation models for SDM involves training large-scale generative Task2ObsSeq models on all available data. Given a task description (e.g. in natural language) the model generates the corresponding observation sequence which can then easily be translated into actions using simpler, agent-specific models. A first example of this approach is demonstrated in the Universal Policy paper. Since large datasets of (task, trajectory)-pairs are challenging to obtain, pooling data across multiple agents (Open X-Embodiment, Octo: An Open-Source Generalist Robot Policy) and leveraging existing video data becomes essential.

Related Projects:

1. Extending the universal policy to agents with differing action spaces:
   
   Making Universal Policies Universal
2. Identifying instructable behaviors from video trajectories in a semi-supervised setting:
   
   Data Augmentation for Instruction Following Policies via Trajectory Segmentation

Foundation Models for Scientific Research

I'm interested in the potential for current foundation models to accelerate scientific research. However, a significant challenge remains in evaluating generated scientific content. This inspired my recent work:

Socials:

Other Interest:

My other interests mostly inlcude sport: 🏄 🏂 ♟ 🥋 🏃 🏊 🧘