Scene understanding is crucial to advance autonomous robotic systems deployment in the real world. The selected student will focus on developing a framework for constructing 3DSGs capable of recording changes in dynamic environments as they occur.

Learning action-conditioned driving dynamics from raw pixels is challenging due to high dimensionality and weak temporal cues. This work combines Latent Action Pretraining (LAPA) with a conditional Diffusion Model to learn discrete latent actions and predict their future evolution. The framework captures multi-modal driving behaviors in latent space, enabling interpretable and data-efficient policy learning for autonomous driving.

In this project, we develop a shared control framework that guarantees safety using control-invariant sets (CISs), which are computed from the robot’s dynamics and an environmental model. The CISs ensure that unsafe human commands are overridden, while safe commands are executed normally.

We are currently seeking for a motivated and talented master’s student to work on discriminative filtering and feature extraction for classification of tactile signals.