The goal of the online BeNeRL seminar series is to invite RL researchers (mostly advanced PhD or early postgraduate) to share their work. In addition, we invite the speakers to briefly share their experience with large-scale deep RL experiments, and their style/approach to get these to work.

Abstract: In visual multi-object manipulation decision-making tasks, understanding object interactions from pixel observations is a foundational challenge for real-world applications of reinforcement learning and robotics. Recent advances in unsupervised object-centric representations offer a promising path, as they decompose the image-based inputs into interpretable structures that naturally align with multi-object manipulation. In this talk, we explore recent advances that utilize object-centric representations for more efficient policy learning. We begin by introducing Deep Latent Particles (DLP), a representation that leverages keypoint-based object discovery to capture spatial and appearance features in a compact, interpretable form. Next, we discuss the application of these representations in model-free reinforcement learning, demonstrating how an Entity-Centric RL approach enables efficient learning for tasks requiring multi-object interaction and the emergence of compositional generalization. Extending these ideas, we present EC-Diffuser, a diffusion-based behavioral cloning framework to learn policies from demonstrations, achieving zero-shot generalization across complex object configurations. Finally, we introduce Deep Dynamic Latent Particles (DDLP), which extends DLP to video prediction, providing a bridge to model-based RL by predicting object dynamics over time.

Bio: Tal Daniel is a final-year Ph.D. student in the Electrical and Computer Engineering faculty at the Technion, where he earned his B.Sc. and M.Sc., under the supervision of Prof. Aviv Tamar, and an incoming postdoc at Carnegie Mellon University (CMU). He is the winner of The Miriam and Aaron Gutwirth Memorial and The Irwin and Joan Jacobs Ph.D fellowships. His research interests include unsupervised representation learning, generative modelling and reinforcement learning.