The Dawn of Domain-Specific Hardware System for Autonomous Machines

Autonomous machines such as drones, robots, and even Augmented/Virtual Reality headsets, while are of  a computing nature, intimately interact with both the environment and humans. They must be built, from the ground  up, with principled considerations of three main components: imaging, computer systems, and human perception.  While our community has been, for very good reasons, focused almost exclusively on improving the computer systems,  our position is that the continued progress in autonomous machine computing (AMC) must rely on co-designing and  co-optimizing all three components. We call them, collectively, agents for progress in AMC. This talk has three goals:  discuss how the three agents play their (largely isolated) roles in today’s AMC, describe a framework where the three  agents are fundamentally connected, and finally, present exciting research opportunities that arise when jointly  designing and optimizing across the three agents. 

Integrated Task and Motion Planning (TAMP) presents a significant computing challenge in robotics,  requiring the seamless coordination of high-level task planning with low-level motion planning while navigating  complex environments. In this work, we propose a novel framework for TAMP based on the brain-inspired neuro vector-symbolic paradigm. Our framework harnesses the inherent properties of vector-symbolic architecture to encode  and manipulate information in a high-dimensional space, enabling efficient representation and manipulation of  complex task and motion planning scenarios. We demonstrate the effectiveness of our approach through experimental  validation in some robotic domains, showcasing its ability to generate feasible and efficient plans in real-world  environments. Our work not only advances the field of TAMP but also contributes to the broader exploration of brain inspired computing paradigms for robotics applications.

The development of autonomous machines has gained significant traction in recent years. However,  ensuring the safety and robustness of these systems remains a critical challenge. This talk presents novel system  architectures and learning algorithms designed to address these challenges and enhance the safety and robustness of  Autonomous Machine Control (AMC) systems. At the hardware level, we propose "asymmetric resilience" coupled  with fault-tolerant methods to cope with transient faults in the system while balancing size, weight, area, power, and  cost metrics. At the software level, we present feedback-directed deep learning methods that continuously assess the  system's state and environment and can intervene if potentially unsafe situations are detected. This system architecture  is designed to be robust to uncertainties, disturbances, and variations in operating conditions. We evaluate the system  through simulations and real-world experiments on AMC systems, and the results show significant improvements in  safety compared to traditional approaches. The proposed safety-aware robust architectures and learning algorithms  represent a much-needed step towards developing AMC systems that can be deployed in real-world settings.

Commercial autonomous machines is a thriving sector, one that is likely the next ubiquitous computing  platform, after Personal Computers (PC), cloud computing, and mobile computing. Nevertheless, a suitable computing  substrate for autonomous machines is missing, and many companies are forced to develop ad hoc computing solutions  that are neither principled nor extensible. By analyzing the demands of autonomous machine computing, this talk proposes Dataflow Accelerator Architecture (DAA), a modern instantiation of the classic dataflow principle, that  matches the characteristics of autonomous machine software.