Repeating self-driving experiments done on Lexus on minicars: Mixture density networks and energy-based models on Donkeycar S1.
Organization
Autonomous driving lab
Abstract
Testing models on the real car is cool but comes with a cost. In a day of testing, we are able to do a maximum of 10 test runs. Moreover, during the day and between days conditions can change and it is hard to guarantee a fair comparison. Also, on real roads, the challenging situations are few and far between - during 25 minutes test run, there are 10-15 challenging situations (sharp turns, intersections).
All this means that if we want to compare 3-4 approaches to self-driving, we need to spend many many days on the test route to have enough evidence to claim the superiority of any approach over another. With the 1:10 scale minicars the testing of solutions is significantly faster and simpler:
- No need to spend an hour driving to the test location and back (+ time saved on cleaning the car from mud)
- One lap on a reasonably-sized route takes up to a minute instead of 25 minutes.
- One can construct a variety of “roads” with any density of difficult situations.
- Weather conditions, including light conditions, are much more stable indoors, allowing fair comparison across time
In 2022 we have worked on comparing different network architectures’ ability to drive the real car based on camera images. This includes neural networks performing regression and classification, mixture density networks, and energy-based models. With a variety of models needing testing, we need to draw conclusions from just 3 runs per model, meaning we lack statistical power. In the present thesis, you would train the same types of neural networks for the minicar and evaluate their abilities using the same metrics and ideas as was done on the real car. Likely the number of repetitions you can run is higher and differences between approaches will be more significant. This work aims to demonstrate that certain studies we can do on the real car can actually be more effectively done on the minicar.
All this means that if we want to compare 3-4 approaches to self-driving, we need to spend many many days on the test route to have enough evidence to claim the superiority of any approach over another. With the 1:10 scale minicars the testing of solutions is significantly faster and simpler:
- No need to spend an hour driving to the test location and back (+ time saved on cleaning the car from mud)
- One lap on a reasonably-sized route takes up to a minute instead of 25 minutes.
- One can construct a variety of “roads” with any density of difficult situations.
- Weather conditions, including light conditions, are much more stable indoors, allowing fair comparison across time
In 2022 we have worked on comparing different network architectures’ ability to drive the real car based on camera images. This includes neural networks performing regression and classification, mixture density networks, and energy-based models. With a variety of models needing testing, we need to draw conclusions from just 3 runs per model, meaning we lack statistical power. In the present thesis, you would train the same types of neural networks for the minicar and evaluate their abilities using the same metrics and ideas as was done on the real car. Likely the number of repetitions you can run is higher and differences between approaches will be more significant. This work aims to demonstrate that certain studies we can do on the real car can actually be more effectively done on the minicar.
Graduation Theses defence year
2022-2023
Supervisor
Ardi Tampuu
Spoken language (s)
Estonian, English
Requirements for candidates
Has passed Machine Learning and Neural networks courses.
Level
Masters
Keywords
Application of contact
Name
ardi tampuu
Phone
E-mail