IPAB Workshop-28/03/2024

Speaker: Cheng Wang

Title: Modelling realistic driving behaviour in simulations for testing autonomous vehicles

Abstract: Autonomous vehicles, defined as safety-critical robotics, are facing the challenge of safety verification and validation for deployment in the real world. Testing autonomous vehicles in the real world is risky, alongside demands on time and financial resources. In contrast, simulation offers a promising alternative to real-world testing, provided that valid simulation environments are established to ensure the credibility of simulation results. To this end, we propose a new paradigm to model realistic driving behaviour in simulations, enabling the creation of realistic driving scenarios for testing autonomous vehicles. In this talk, I will first introduce the requirements for driving behaviour modelling. Following this, I will present the concept we are developing and outline our future work.

Speaker: Yihe Lu

Title: Visuomnemo-tropotaxis: two-sided views along the way

Abstract: Insects can move (taxis) in a desired direction by balancing sensory inputs on left and right sides (tropotaxis); in particular, vision can be used (visuo-tropotaxis), e.g., moving along a corridor using optic flow. On a separate note, some insects can retrace visually familiar routes based on memories of panoramic views (visuo-mnemotaxis). Can these two seemingly different behaviours share similar underlying mechanisms? More specifically, is visuomnemo-tropotaxis useful in navigation? In this talk, I will brief my ongoing work that probes this ability of an insect-inspired model. The model has been tested on differential wheeled robots in visual teach-and-repeat route following tasks.

Speaker: Octave Mariotti

Title: Improving Semantic Correspondence with Spherical Maps

Abstract: Recent progress in self-supervised representation learning has resulted in models that are capable of extracting image features that are not only effective at encoding image level, but also pixel-level, semantics. These features have been shown to be effective for dense visual semantic correspondence estimation, even outperforming fully-supervised methods. Nevertheless, current self-supervised approaches still fail in the presence of challenging image characteristics such as symmetries and repeated parts. I this short talk, I will present our attempt to address these limitations by supplementing discriminative self-supervised features with 3D understanding using a weak geometric spherical prior. Additionally, I will discuss a new evaluation metric that better accounts for repeated part and symmetry-induced mistakes.