Uncertainty-Aware Human Intervention for Autonomous Vehicles

Uncertainty-Aware Human Intervention for Autonomous Vehicles

Key Outcomes

16x Fewer collisions
12x Reduction in computation time
93% Reduction in human takeovers

Context

The Autonomous Driving market is expected to generate between 300 and 400B by 2035 with studies showing it could reduce the number of accidents by 15% by 2030. In 2021, two-thirds of consumers reported interest in purchasing L4 highway pilots for a one time fee of 10K. However, only 4% of cars are projected to include L3+ AD functions by 2030. Moreover, trust in the technology has declined by 10% with 26% of consumers reporting they would switch to AV in 2021, compared to 35% in 2020. Additionally, perception, prediction, and planning algorithms continue to require significant computational resources and are considered to be the remaining areas of high difficulty according to survey respondents.

Indeed, current AV computations would result in carbon emissions equivalent to that of all existing data centers as shown in a recent MIT study.Despite these challenges, several AV products have begun receiving ISO 26262 certifications, e.g., NVIDIA OS, Mobileye, Apex.ai, leaving computational efficiency, consumer trust, and algorithm robustness as the final barriers for large-scale adoption.

Our Solution

The technologies developed by Themis AI have been successfully used to address these limitations. The approaches have been comprehensively tested and demonstrated through over five years of experiments with full-scale vehicles. Research results show that integrating these proprietary uncertainty estimation algorithms with state- of-the-art AVs led to 16x fewer collisions, a 12x reduction in computation time, a 89% success rate when recovering from near-crash scenarios, and a 93% reduction in automated requests for humans to take over the wheel. These algorithms are now part of Themis AI software solutions that can be seamlessly integrated into existing systems and used for applications beyond autonomous driving. More details and additional results have been presented in several academic publications.

Additional Resources

Z. Liu, A. Amini, S. Zhu, S. Karaman, S. Han, and D. Rus, “Efficient and Robust LiDAR-Based End-to-End Navigation,” presented at the 2021 IEEE International Conference on Robotics and Automation (ICRA).

A. Amini et al., “VISTA 2.0: An Open, Data-driven Simulator for Multimodal Sensing and Policy Learning for Autonomous Vehicles,” presented at the 2022 IEEE International Conference on Robotics and Automation (ICRA).

Z. Liu et al., “BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird’s-Eye View Representation,” presented at the 2023 IEEE International Conference on Robotics and Automation (ICRA).

A. Amini et al., “Learning Robust Control Policies for End-to-End Autonomous Driving From Data-Driven Simulation,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 1143–1150, Apr. 2020.

Amini, Alexander, Wilko Schwarting, Guy Rosman, Brandon Araki, Sertac Karaman, and Daniela Rus. “Variational Autoencoder for End-To-End Control of Autonomous Driving with Novelty Detection and Training De-Biasing,” presented at the 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).