1 of 8

Differentiable Autonomous Driving

Xijun Wang

10/21/2021

2 of 8

Motivations and background

[1] Raquel Urtasun - A future with affordable Self-driving vehicles

3 of 8

Motivations and background | Problems

[1] Raquel Urtasun - A future with affordable Self-driving vehicles

Problems:

  • Small interface between tasks 🡪 Error pass and accumulate;
  • The Computation is not shared between modules;
  • Each submodule is trained independently to optimize a different objective

Classical Stack

4 of 8

Motivations and background

[1] Raquel Urtasun - A future with affordable Self-driving vehicles

  • Input: Sensor data;
  • Output: Waypoints, or Steering angle and Acceleration;
  • Advantages: Easy and simple, Just need a high-enough capacity neural net;
  • Disadvantages: lack of interpretability;

​

End-to-End Stack

5 of 8

State-of-art

[1] Prakash, A., Chitta, K., Geiger, A.: Multi-modal fusion transformer for end-to-end autonomous driving. In: CVPR (2021)

6 of 8

Design

  • Early Fusion;
  • Saving up to 50% computational cost compared with muti-model methods;
  • Interpretable: with Classification/Detection outputs;

Self-adapting

Conv

Self-adapting

Conv

Self-adapting

Conv

Self-adapting

Conv

Self-adapting Conv

Encoder

MLP

Loc1

Cls

Loc2

Loc3

Det

Auxiliary

Future trajectory

7 of 8

Goals

  • Short-term
    • Make the proposed framework work normal and settled; (by Nov. 20, 2021)
    • The Performance can reach the SOTA; (by Dec. 10, 2021)
  • Long-term
    • Publish a paper；

8 of 8

Q&A

THANK YOU