Conversation

Alex Kendall
@alexgkendall
No two experiences on our roads are the same. That's why it is so important for autonomous driving systems to be able to learn behaviours and apply them to new scenarios. This is called generalisation. Today we're sharing a single AI model which can generalise to new vehicles:
Embedded video
GIF
3
57
Alex Kendall
@alexgkendall
Watch our van (bottom) following our passenger car (top) driving autonomously through London. Both vehicles are being driven by the same end-to-end neural network with exactly the same neural network weights. This has never been done before.
Embedded video
0:52
491 views
1
19
Alex Kendall
@alexgkendall
This "mini-documentary" tells our story building AV2.0: autonomy which can scale to 100 cities. last year we showed we could generalise to new cities (training in London, deploying in other cities with no new data or HD map) and now to new vehicles:
youtube.com
Wayve Presents: The Industry's First Generalisable AI Driver
The inside story of the Wayve team who created the first AI Driver that can generalise to different cities and vehicle types, a grand challenge of self-drivi...
8
Likes
Alex Kendall
@alexgkendall
Replying to
@alexgkendall
The most exciting bit for me was seeing our van learn more complex manoeuvres in the first week of operation (eg. overtaking double parked buses) which typically only emerges after 1000s hours training. This shows transferring learned behaviour from our passenger vehicle data.
1
3
Alex Kendall
@alexgkendall
Of course we tested our AI extensively in simulation before putting it on the road to ensure it was safe and learning the correct behaviours.
Embedded video
GIF
1
2
Alex Kendall
@alexgkendall
We saw a very common machine learning result: multitask training improves the performance of both tasks individually. This is an important thesis of AV2.0, we can use experience from each new deployment to improve performance in general.
Image
1
8