What data do we need for training an AV motion planner?

Long Chen, Lukas Platinsky, Stefanie Speichert, Błażej Osiński, Oliver Scheel, Yawei Ye, Hugo Grimmett, Luca del Pero, and Peter Ondruska

ICRA 2021

Video

Data are key for ML planning

Planning performance improves with the amount of training data
Training data can be collected in different quantities and qualities from low-cost camera sensors to high-cost lidar+camera+radar configurations.

Collecting large-scale ML planning datasets using fleet

Balancing the trade-offs we can collect large-scale datasets using low-cost sensors to train ML planning systems deployed on SDVs with expensive high-end sensors.

Finding the optimal hardware configuration

We developed a pipeline to simulate lower-quality data from low-cost sensors, and evaluate the ML planner performance, in order to find the optimal sensor configuration

Simulated lower-quality data is create, by limiting the range and FoV, or changing the perception accuracy by adding random noises to the agent positions

Impact on ML Planner Performance

Sensitivity Analysis on Data Quality

ML Planner needs at least 40m range, 130 FoV to have acceptable performance, and is not sensitive to rotational error.

Quality vs Quantity

With domain adaptation, greater amount of lower quality data beat small amount of high-quality data

Cite

@inproceedings{chen2021data,

title={What data do we need for training an AV motion planner?},

author={Chen, Long and Platinsky, Lukas and Speichert, Stefanie and Osinski, Blazej and Scheel, Oliver and Ye, Yawei and Grimmett, Hugo and Del Pero, Luca and Ondruska, Peter},

booktitle={2021 IEEE International Conference on Robotics and Automation (ICRA)},

pages={--},

year={2021},

organization={IEEE}

}