Safedrive Dreamer: Navigating Safety-Critical Scenarios in the Real-world with World Models
Keywords: Safedrive Dreamer, Autonomous Driving (AD), World Models, Safety-Critical Scenarios, Sim-to-Real Transfer
TL;DR: We proposed "Safedrive Dreamer", a novel vision-based navigation framework that integrates world models with safety-critical decision ability, enabling autonomous vehicles to navigate complex situations safely in the real world.
Abstract: Ensuring safety in dynamic and unpredictable environments is a crucial challenge in the rapidly evolving field of autonomous driving. In this work, we propose the Safedrive Dreamer, a novel vision-based navigation framework that integrates world models with safety-critical decision ability, enabling autonomous vehicles to navigate complex situations safely in the real world. Our approach proactively learns potential dangers and plans safer routes, leveraging the predictive capabilities of world models and significantly reducing the reliance on extensive trial-and-error learning in the real world. The effectiveness of Safedrive Dreamer is validated through a series of experiments in real-world sim-to-real driving conditions, covering a diverse range of safety-critical scenarios, such as abrupt obstacle avoidance. Our results show that Safedrive Dreamer achieves superior performance in safety metrics, such as collision avoidance and risk minimization, compared to other end-to-end solutions. This framework advances autonomous driving safety and offers insights into integrating world models for enhancing decision-making in safety-critical applications. Safedrive Dreamer paves the way for developing more resilient and trustworthy autonomous driving systems that are adept at handling the dynamics and uncertainties of the real world.
Submission Number: 14
Loading