Workshop on Autonomous Driving

Event: CVPR 2025 Workshop on Autonomous Driving · Duration: 30 min · ▶ Watch on YouTube

Abstract

This talk provides an overview of trends in AI-based robotics, focusing on probabilistic robotics, deep learning, and foundation models, and their application to robot navigation and autonomous driving. It highlights the success of probabilistic methods in state and uncertainty estimation for robust systems, and the role of deep networks in perception and prediction. The speaker addresses the challenges posed by high-definition maps in dynamic environments and proposes leveraging uncertainty-aware panoptic segmentation and foundation models to build more robust and flexible autonomous systems, moving towards embodied AI.

Speakers

  • Wolfram Burgard — University of Technology Nuremberg

Talks (1)

  • 00:04:00 — Wolfram Burgard: Probabilistic and Deep Learning Approaches to Robot Navigation and Autonomous Driving
    • This talk discusses the evolution of AI-based robotics, from probabilistic methods to deep learning and foundation models, emphasizing their application in autonomous driving and the challenges of high-definition maps.

Key Takeaways

  • Probabilistic robotics has been foundational for robust state and uncertainty estimation in autonomous systems, particularly for localization and mapping.
  • Deep learning significantly enhances perception and prediction capabilities, but traditional HD maps present challenges in dynamic environments due to acquisition, maintenance, and generalization issues.
  • Uncertainty-aware panoptic segmentation (for both images and LiDAR) is crucial for integrating deep learning outputs into probabilistic frameworks, providing calibrated uncertainties for downstream tasks like localization.
  • Foundation models offer a promising path towards more generalized and transferable robotic capabilities, enabling language-guided navigation and reducing reliance on explicit, static maps.
  • Building truly embodied AI systems for robotics and autonomous driving requires a synergistic combination of probabilistic robotics, deep learning, calibrated uncertainties, and the development of specialized robotics foundation models.

Methods / Models / Datasets Mentioned

  • KUKA omniMove
  • Pose-Graph-SLAM
  • dotscene
  • EvPSNet
  • EvLPSNet
  • uPLAM
  • LaneGraphNet
  • Graph-RCNN
  • LaneDirNet
  • HOVSG (Hierarchical Open-Vocabulary 3D Scene Graphs)
  • FM-Loc
  • NetVLAD
  • PatchNetVLAD
  • Open X-Embodiment
  • PhysBench
  • GPT-4o
  • Gemini-1.5-pro
  • GPT-4o-mini
  • GPT-4V
  • LLaVA-interleave
  • Phi-3V
  • VILA-1.5
  • LLaVA-NeXT-Video

Topics

Autonomous Driving · Robot Navigation · Probabilistic Robotics · Deep Learning · Foundation Models · Semantic Segmentation · Panoptic Segmentation · Uncertainty Estimation · Localization · Mapping · Language-Guided Navigation

Notes

Open for commentary — connections to other work, critiques, follow-up reading.