End-to-end Driving
Tambet Matiisen
7.09.2022
Modular Approach
�Sensors�
�Perception�
�Planning�
�Control�
�Actuators�
camera image
detected objects
trajectory
steering, �gas and brake
End-to-End Approach
�Sensors�
�Neural Network�
�Actuators�
camera image
steering, �gas and brake
Two approaches
| Modular Approach | End-to-End Approach |
Pros |
|
|
Cons |
|
|
Tesla
comma.ai
Agenda
Agenda
Agenda
Imitation learning
NVIDIA DAVE-2
Problem: distribution drift
Distribution drift example
Common solutions to distribution drift
Common solutions to distribution drift
Common solutions to distribution drift
Common solutions to distribution drift
Shift and rotate
Common solutions to distribution drift
DAGGER algorithm:
Common solutions to distribution drift
Practical DAGGER algorithm:
DAGGER
Imitation learning summary
Agenda
Reinforcement learning
Learning to drive in a day
Simulations
Reinforcement learning is usually done in simulations, because who would want to crash the car in the real world?
Microsoft
AirSim
CARLA
Distribution mismatch
Simulations do not really match the real world:
Unsupervised domain adaptation
Wayve sim2real
Reinforcement learning summary
Agenda
Agenda
Network inputs
Network inputs
ALVINN (1995)
Network inputs
ALVINN (1995)
Network inputs
NVIDIA DAVE-2 (2016)
Network inputs
NVIDIA DAVE-2 (2016)
Network inputs
DAVE (2005)
Network inputs
Hecker et al. (2018)
Network inputs
Müller et al. (2018)
Network inputs
Codevilla et al. (2019)
Network inputs
Codevilla et al. (2017)
Network inputs
Codevilla et al. (2017)
Network inputs
Codevilla et al. (2017)
Network inputs
Hecker et al. (2019)
Network inputs
Hecker et al. (2019)
Network inputs
Zeng et al. (2019)
Network inputs
Bansal et al. (2018)
Network inputs
Multi-modal fusion
Multiple timesteps
Network inputs summary
Agenda
Network outputs
Network outputs
Pros:
Cons:
Network outputs
Pros:
Cons:
Network outputs
Curvature is the inverse of turning radius.
Network outputs
Pros:
Cons:
Network outputs
Network outputs
Pros:
Cons:
Network outputs
Multiple equally good trajectories:
Network outputs
Network outputs
Pros:
Cons:
Network outputs
Affordances represent semantic information used by simple planner to plan trajectory, for example:
Network outputs
Pros:
Cons:
Network outputs summary
Agenda
Evaluation methods
Open-loop evaluation
Model predictions are compared with human driver ground truth values, e.g. steering wheel angle and speed.
Closed-loop evaluation
The model drives the car, some metric is used to measure driving ability, e.g. kilometers driven without intervention.
BAD!�But could be used in model architecture search phase.
GOOD!�But beware of different driving conditions!
Closed-loop evaluation metrics
State of California disengagement report 2021
Evaluation summary
Agenda
Interpretability
Neural networks are generally considered black boxes. If the network makes an error, it may be hard to understand why it happened and how to fix it.
Highlighting salient areas on the image
Several methods can be used to highlight the areas on the image that were used to make the decision:
Auxiliary outputs
Forcing the network to predict additional outputs, e.g. semantic segmentation, can both speed up training and make the model generalize better, due to more supervision signal. These outputs can be used at prediction time for debugging.
Auxiliary outputs
Interpretability summary
Agenda
Is End-to-End the Future?
Putting stickers on road misguided Tesla Autopilot
Showing this image to Tesla activated windscreen wipers
Adding stickers to stop sign caused it to be misclassified
Ekholt et al., Robust Physical-World Attacks on Deep Learning Models (2017)
Safety summary
Agenda
Company: Tesla
Network inputs:
Network outputs:
Company: comma.ai
Network inputs:
Network outputs:
Company: Wayve
Network inputs:
Network outputs:
Companies summary
Thank you!
tambet.matiisen@ut.ee