Grant: Discovery Grants Program - Individual

Robots are gradually evolving out of orderly manufacturing facilities into households, executing simple tasks such as vacuum cleaning and lawn mowing. Meanwhile, by ever-advancing artificial intelligence (AI), robots are expected to go beyond this by performing complex interactions with humans and the physical world. But this will not happen until robots are equipped with spatial awareness, beyond reporting 'what' is 'where' in an image. The research area of simultaneous localization and mapping (SLAM) in robotics and computer vision provides a geometric understanding of the environment. SLAM is a process in which an unknown environment is explored and mapped consistently by a sensor, allowing a robot to determine its position while understanding the geometry of that environment. SLAM is increasingly evolving towards dense and semantic perception, aiming to create a human-level understanding of the environment. Thus, the development of versatile and advanced SLAM algorithms, referred to as 'spatial AI', will have a great impact on robotics, and particularly in areas such as autonomous cars, visual Internet of Things, virtual reality, and augmented reality. This research proposal concentrates on spatial AI, targeting four areas: 1) energy consumption, 2) run-time speed, 3) robustness, and 4) versatility. For a robotic perception system, the set target for energy consumption is a battery charge once a day (less than 1 Watt power usage), and the set target for frame-rate is 3000 FPS. At this frame-rate, 1 cm self-motion at 110 km/h is perceived. Moreover, a robust and reliable system is crucial when robots are working with humans. Finally, a versatile system is needed to infer objects' properties such as position and orientation, geometric shape, colour, texture, thickness, weight, etc., all from limited incoming data. This work is based on the hypothesis that machine learning algorithms will help with versatility and robustness, and mixed analog-digital computation will help in achieving the desired energy consumption and run-time speed. This proposal presents a unique and novel scheme for visual processing pipelines for robotics and real-time applications where hardware and software are co-designed. In robotics, understanding the physical environment and determining the motion of the robot is the cornerstone of autonomy. A robotic arm in a farm, a self-driving car on a road, a mobile robot in a hospital, and an autonomous submarine in the deep ocean all need to have an accurate perception of the environment and to know their location within the environment to accomplish their tasks. The Canadian autonomous car and robotics industries will benefit by employing the outcomes of the proposal, i.e. efficient perception systems in terms of functionality and resource utilization. Moreover, emerging and applied research fields, e.g. applied perception and robotics in health and the environment, will have a reliable platform to build their solutions on.