Melvin's digital garden

Autonomous Vehicles: advances, algorithms, and the road ahead

[2016-09-30 Fri 18:59:34] speaker: Benjamin Choi, DSO National Labs event: Friday Hacks

1950s - 1970s: smart roads

1980s: Autonomous Land Vehicle project

  • road following

1990s: No hands across America

Early 2000s: PerceptOR

2004-2005: DARPA Grand Challenge

2007: DARPA Urban Challenge

2009-now:

  • Google gets in the game
  • Mining trucks are self driving
  • nuTonomy and Uber
  • George Hotz making a kit
  • DSO working on autonomous armored vehicles

Why is it hard?

  • roads are dynamic
  • reading traffic lights
  • driving in the rain
  • offroad jungle paths
  • no GPS signals

offroad challenges

  • paths include soft objects which the vehicle can bash through
  • LIDA can only detect surfaces, but not tell if it is hard or soft
  • perhaps a camera can help, but computer vision is hard
    • illustrate with grid of numbers
  • paths are not flat, need to determine if safe to drive over

subsystems for autonomy

  • localisation
    • odometry by tracking points on images
    • LOAM algorithm
  • perception
    • semantic labelling of image
      • Efficient graph-based image segmentation
    • transform coordinates and fuse with LIDA data
    • dynamic obstacle tracking for moving objects
    • using neural networks to analyze images
  • planning
    • need to account for vehicle dynamics
    • A* search on state lattice
  • control
    • open loop control, no feedback
    • feedback control

other issues

  • missing planning
  • individual and cooperative behaviours
  • mapping
  • visualization

engineering robustness

  • health reporting
  • supervisor daemons
  • state machines for each module
  • safeguards
  • deterministic execution times
    • custom dafe data structures

the road ahead

  • full autonomy
  • deep neural networks
  • solid state LIDAR
  • toolkits

questions

  • testing in the lab
    • with simulators
  • AVs are getting bullied

Links to this note