Abstract
dc:descriptionAutonomous robots have great potential to improve our day-to-day lives. Hence, their demand for real-time and safety-critical tasks is increasing, including medical care, home assistance, and autonomous driving. An autonomous robot system consists of several latency-sensitive computation tasks. Further, a crucial aspect of computing for autonomous robots is energy efficiency, as computation can contribute to a significant fraction of the total power consumption. With the recent advancements in machine learning, algorithms for autonomous robots’ operation are explored extensively at the software level. However, their architectural implications are not explored in detail. This thesis explores algorithm and hardware design for energy-efficient computation acceleration of perception and planning tasks in autonomous robotics. It first proposes a motion planning hardware accelerator, MPAccel. MPAccel consists of a Spatially-Aware Scheduler and Collision Prediction Units. The proposed scheduler and predictor exploit the physical-spatial locality of a robot's positions in the environment to reduce redundant computation. Further, MPAccel proposes a hierarchical collision detection approach and corresponding hardware accelerator. MPAccel enables real-time sampling-based motion planning for a 7-DOF robotic arm with less than 1ms latency and 3.5W power consumption. The second part of the thesis focuses on regression networks, as they are used in autonomous robot perception and planning. It explores the design space of regression by binary classification and proposes Binary-Encoded Labels (BEL). This thesis further analyzes the properties of suitable label encodings and proposes manually designed label encodings. It also proposes an end-to-end training approach to learn label encodings and network parameters together for a given task. BEL reduces regression error by 10.9% compared to direct regression, enabling the use of smaller and/or sparser networks for compute- and energy-efficient execution. Further, BEL-based regression networks are explored for smaller and sparser neural motion planners, reducing neural planning computation by a factor of 11.4x. The third part of the thesis explores the impact of soft errors on collision detection hardware accelerators in motion planning. It proposes an application-specific reliability metric, Collision Exposure Factor, and demonstrates its use for faster fault characterization and low-overhead selective error mitigation.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Electrical and Computer Engineering
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shah, Deval
- Contributors dc:contributor
-
- Aamodt, Tor M.
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/88403