SafetyNet: Safe planning for real-world self-driving vehicles using machine-learned policies

Matt Vitelli, Yan Chang, Yawei Ye, Maciej Wołczyk, Błażej Osiński, Moritz Niendorf, Hugo Grimmett, Qiangui Huang, Ashesh Jain, Peter Ondruska

Video

Merging ML and expert planning

  • ML planning is a scalable way for self-driving.

  • ML planning can not offer safety guarantees.

  • Hybrid ML planning scales with data while assuring the safety.

Hybrid ML planning system

System overview: SafetyNet combines a machine-learned planner with an effective rule-based fallback layer to deliver safe learned planning for SDVs.

Model architecture: Our ML planner model is built on GNNs consist of a PointNet-based local subgraph and a global graph using a Transformer encoder.

Model Training: Our model is trained using imitation learning on 300hr training data collected at Palo Alto and San Francisco.

Real world driving results

The results in this section show the SafetyNet controlled SDVs driving in downtown San Francisco.

Stopping for a Lead Vehicle

Stopping for Red Traffic Light

Adjusting Speed for a Vehicle Cut-In

Slowing for a Perpendicular Cut-In

Nudging around a Vehicle on the Left

Nudging around a Parklet on the Right

Nudging around a Truck on the Right

Safely Driving through a Yellow Light

Impact of SafetyNet

Safety events as a function of dataset size for ML planner with- and without fallback layer. We see that the addition of fallback layer significantly improves safety metrics.

Compared with ML planner-only approach, SafetyNet significantly reduces collisions, close calls, and discomfort braking, at the expense of slightly more passiveness.


Cite

@article{vitelli2021safetynet,

title={SafetyNet: Safe planning for real-world self-driving vehicles using machine-learned policies},

author={Matt Vitelli, Yan Chang, Yawei Ye, Maciej Wołczyk, Błażej Osiński, Moritz Niendorf, Hugo Grimmett, Qiangui Huang, Ashesh Jain, Peter Ondruska},

year={2021},

eprint={2109.13602},

archivePrefix={arXiv},

}