Search Articles

Terminology #

• Robots - Machines the put compute in motion

Overview #

• Machines can be
  • Descriptive - Explaining what is happening
  • Predictive - Predicting what will happen
  • Prescriptive - Prescribing what should happen and determining what it should do
• Robots require a body (hardware) and a brain (software)
• There are three types of learning for robots (similar to other ML models)
  • Supervised learning - Learning from labeled data
  • Reinforcement learning - Learning from rewards RL
  • Unsupervised learning - Learning from unlabeled data

Decision Making for Robots #

• This contains three stages
  • Perception - The robot needs to perceive the world around it
    • This is done through sensors
    • A CNN is used to do image segmentation (label different parts of the image as certain classifications)
      • State of the art is like 90% – this is not good enough for robotics (e.g. cars)
    • There are a lot of potential attacks like adversarial perturbations that you need to be aware of
  • Planning - The robot needs to plan what it should do
    • This is done through a model of the world
    • This is the brain of the robot
  • Action - The robot needs to act on the world
    • This is done through actuators
    • This is the output of the robot
• Reinforcement learning RL is used for the planning and action stage
  • This differs from supervised learning because you don’t need input/output pairs

Fully Autonomous Vehicles #

• Lidar made a huge difference in enabling autonomous vehicles
  • Before we only had sonar which wouldn’t work
• Sensors don’t work well in weather
  • The uncertainty of the sensors increases greatly of rain or snow
• Uncertainty also increases hugely in congested areas, with other objects on or near the road, etc.

The Traditional A/V pipeline

1. Process sensor data
2. Detect obstacles
3. Localize the car relative to the other objects
4. Plan a path
5. Actuate the car

How do you connect these modules together – through what parameters?

• Using a large dataset we can learn an underlying representation of driving based on how humans drive in similar situations
• How can we use machine learning to go directly from perception to actuation (cutting out all of the middle parts)
• Can train a deep neural net to go from the raw perception input data to actuation
  • Input:
    • Sensors: Camera, lidar, etc.
    • Coarse Maps: GPS
• We need a lot of data to train this model
  • We can generate this in simulation, see Vista and Vista Noes

Autonomous Recipe - Making your car autonomous #

1. Make car drive by wire
2. Extend the car with cameras, lidars
3. Use existing software modules
   • A perception module
   • An estimation/learning module
   • A planning/control module