Learning-based offline & online auto-calibration of autonomous vehicle controls- ON-305

Desired discipline(s): Engineering - computer / electrical, Engineering, Computer science, Mathematical Sciences
Company: Gatik
Project Length: 4 to 6 months
Preferred start date: As soon as possible.
Language requirement: English
Location(s): ON, Canada
No. of positions: 1
Search across Mitacs’ international networks - check this box if you’d also like to receive profiles of researchers based outside of Canada: 

About the company: 

Gatik was founded in 2017 by veterans of the autonomous technology space. Gatik’s mission is to deliver goods safely and efficiently using autonomous (self-driving) vehicles. The company focuses on business to business (B2B), short-haul logistics for the retail industry.

Light-duty trucks and vans equipped with Gatik’s proprietary autonomous driving technology move goods on fixed, repeatable routes between stores, distribution centres and collection points, known as the “middle mile” - the most expensive and challenging part of the supply chain for retailers to contend with. It’s also a hugely underserved segment of the autonomous vehicle market, with many others focusing on sidewalk robots or long-haul trucking.

Gatik’s autonomous vehicles have been operating in North America with multiple Fortune 500 retail partners including Walmart since July 2019. The partnership is a first in the AV industry: this kind of middle-mile integration with a major retailer has never been done before. It marked the first-ever deployment of a hub-and-spoke autonomous vehicle delivery model, and is proof of commercialization and scalability for our use case and market.

As a result of COVID-19, Gatik’s autonomous delivery solution is more important than ever. Gatik’s business model and product offerings enhance operational efficiency, public health and community safety via on-road automation and contactless delivery. Specifically, Gatik is helping to lead pandemic response efforts in Canada in the following ways:

  • Reducing human-to-human transmission channels of COVID-19 and other communicable diseases via contactless delivery
  • Minimizing personnel-based disruptions to the supply chain via safe, automated and sustainable operations
  • Empowering communities by delivering goods to local, reliable, convenient pick up points
  • Delivering essential goods in multi-temperature compartments (fresh and frozen food, pharmaceutical, medical equipment)

Please describe the project.: 

Humans have the ability to quickly learn the required throttle or brake to apply to achieve desired acceleration throughout the driving process. These learned abilities allow humans to quickly adapt this throttle or brake mapping to dynamically changing environmental conditions such as a change in payload or a change in weather.

However, this intuitive mapping from throttle/brake to a given acceleration is not as easily established in Autonomous systems today. Autonomous vehicles rely on experimentally determined calibration tables (Hua Huang 2016) or a set of functions carefully developed to approximately mimic the natural mapping. Such approaches are largely labor-intensive and moreover do not account for any external environmental constraints.

The goal of this research is to develop a Learning-based method that will learn the initial throttle/brake mapping and then adapt this mapping when environmental conditions change. To accomplish this there are two stages to the learning process. The offline stage learns the initial mapping from human demonstrations and the online stage continues to modulate the mapping as environmental constraints change.

(Fan Zhu 2008) had proposed a similar approach where human data is used to learn a mapping from throttle to desired acceleration. However, the models suggested by (Fan Zhu 2008) fail to utilize environment variables as inputs to the model reducing their ability to react to these dynamic changes.

This work is targeted at expanding that model both by including environmental constraints such as weather and payload and by relying on more recent developments in machine learning to improve the learning algorithms.

Success of this project, accelerates the development and deployment of autonomous vehicles within urban environments essential for online order fulfillment and contactless deliveries.

Required expertise/skills: 

  • Experience of designing and/or training Neural Networks (NNs)
  • Excellent knowledge of Linear Algebra, Convex Optimization, Statistics or Numerical Analysis
  • Experience in vehicle dynamics, specifically handling/ride performance
  • Expert in analyzing vehicle performance via system modeling and data analysis
  • Experience in active suspension systems, brake systems (brake traction control or stability control), powertrain traction control, electric power steering, or other chassis control systems
  • A solid foundation of control theory
  • Experience with C/C++, Python
  • Worked with one or more neural network frameworks, such as Tensorflow or Torch
  • Ability to both prototype as well as ship machine learning products
  • Experience in software design for safety-critical systems
  • Experience in or around automated driving applications