Robust perception across diverse conditions and extended ranges is critical for L2+/L3 ADAS and essential for safe decision-making in L4 autonomous trucking. Achieving this requires moving beyond single-sensor pipelines toward tightly integrated, multimodal spatiotemporal scene understanding.
In this talk, we present Torc’s AV 3.0 approach to building robust perception systems through unified, end-to-end differentiable spatiotemporal representations. We will discuss key challenges in real-world deployment, including adverse conditions, as well as system latency and compute constraints, and how these shape model and system design under strict long-range perception requirements. We will also cover key training considerations, including scalable learning and continuous improvement, to meet the performance and safety goals required for automated and autonomous driving.
