People of ACM - Shaoshan Liu
April 19, 2022
Why did you decide to make autonomous vehicles and distributed systems a focus of your career?
As an increasing number and kinds of autonomous machines enter our daily life, the age of autonomous machines is upon us and a whole new era of information technology begins. Obviously, autonomous machines have the potential to completely upend our daily life and our economy in the coming decade. Yet, autonomous machines are extremely complex systems that integrate many pieces of technologies and for autonomous machines to become an integral part of our daily life, we are still facing many technical and non-technical challenges, on which I have provided the details in an article titled “Rise of the Autonomous Machines.” To me, R&D of autonomous machines, including robotics, autonomous vehicles etc. is the most exciting area to work on today.
However, autonomous computing systems, including on-machine computing systems, autonomous machine cloud computing systems, and cooperative autonomous machine computing systems are essential to empower autonomous machines. In a way, we now have a golden opportunity to re-design the whole computing stack, from hardware to software, and from mobile to cloud, for autonomous machine workloads. As a computer scientist, nothing is more exciting and challenging than working on autonomous machine computing today.
What is Autonomous Machine Computing?
The field of 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.
As opposed to other computing workloads, autonomous machines have a very deep processing pipeline, or computation graph, with strong dependencies between the different stages and strict deadlines associated with each stage. This characteristic of autonomous machine computing calls for a new architectural model, which we call the Dataflow Accelerator Architecture (DAA). DAA has two ingredients. First, it incorporates a diverse set of domain specific accelerators to address the increasing performance requirement of ever more complicated algorithms. Second, DAA organizes accelerators in a dataflow fashion to remove the inefficiency of centralized coordination and communication. The dataflow organization of accelerators translates the orders of magnitude efficiency gains of individual accelerators to the end-to-end application.
In addition to developing computer hardware for autonomous machine computing, we believe that software optimization has not been fully exploited and there is still plenty of room for improvement. For instance, we have demonstrated that compression-compiler co-design can enable real-time autonomous driving workloads on low-end System-on-Chips (SoCs), confirming that a large potential is yet left untapped in enabling real-time artificial intelligence (AI) on mainstream end devices.
What is the key visual perception problem for autonomous vehicles right now and what is PerceptIn’s contribution to overcoming this?
I believe that vision is the most natural way of perception, and it provides a rich amount of detailed information required for the autonomous machines to understand their environments. The key technical challenge for PerceptIn is sensing-computing co-design to make our product more efficient and affordable.
Specifically, through commercial deployments, we have thoroughly studied the characteristics of real-world autonomous machine workloads and provided key insights on why existing commercial SoCs are not optimized for autonomous machine computing. Due to the lack of suitable computing systems, many companies chose hardware overprovision, resulting in high cost and energy consumption.
Based on the understanding of the autonomous machine workloads, we have developed proprietary autonomous machine computing systems to enable our products. The enabling technology consists of specialized hardware, Sensing-Computing Co-Design methodology, and reconfigurable dataflow accelerator architecture. These innovations addressed the key shortcomings of commercial SoCs and laid the foundation of autonomous machine computing.
However, it was not an easy journey for us, as we have hit a lot of roadblocks along the way. Thus we summarized our learnings of developing autonomous machine computing systems in this article, hoping to provide more business insights to the autonomous machine startup community.
While many cars now have reached partial driving automation (often classified as ‘Level 2’) in which a car can steer, accelerate, or decelerate so long as a human sits in the driver’s seat, full automation (often classified as ‘Level 5’), in which a car can go anywhere or do anything without the need of human supervision, is seen by many as decades away. In your view, how far away are we from having many fully autonomous vehicles on our roads?
I was asked the same question when I started PerceptIn in 2016, and I gave a very pessimistic answer, 2040. Today, I still believe that a universal deployment of fully autonomous vehicles will not happen until 2040. A universal deployment of fully autonomous vehicles is not merely a technical question, it is highly related to the economic development status in each country, the maturity of the autonomous driving supply chain, the regulatory landscape, etc.
For instance, as a startup company in the autonomous driving space, we have undergone four years of painful experiences dealing with a broad spectrum of regulatory requirements. Compared to the software industry norm, which spends 13% of their overall budget on compliances, we were forced to spend 42% of our budget on compliances. Our situation is not alone and, in a way, reflects the dilemma of the artificial intelligence (AI) regulatory landscape. I believe the root cause is the lack of AI expertise in the legislative and executive branches, leading to a lack of standardization for the industry to follow.
In what unexpected ways might fully autonomous vehicles transform society?
One exciting application for autonomous vehicles is healthcare, as we believe that Autonomous Mobile Clinics, or the utilization of autonomous vehicles for healthcare services delivery, is a revolutionary and effective way of healthcare service delivery.
We are facing a global healthcare crisis today and healthcare costs continue to climb, but with the aging population, government fiscal revenue is decreasing. To create a more efficient and effective healthcare system, three technical challenges immediately present themselves: healthcare access, healthcare equity, and healthcare efficiency. An autonomous mobile clinic solves the healthcare access problem by bringing healthcare services to the patient at the order of the patient’s fingertips.
Especially with the current COVID-19 situation, autonomous mobile clinics provide a perfect environment for Point of Care Test (POCT), which refers to multiple clinical tests carried out at the point of care near the patients, rather than from clinical laboratory. When dealing with patients with infectious diseases, lab tests such as for White Blood Cells and C-reactive Proteins are of the utmost importance. Autonomous mobile clinics provide an isolated environment for POCT at the location where the patients reside, effectively constraining the spread of infectious diseases.
Nevertheless, to enable a universal autonomous mobile clinic network, a three-stage technical roadmap needs to be achieved. In stage one, we focus on solving the inequity challenge in the existing healthcare system by combining autonomous mobility and telemedicine. In stage two, we develop an AI doctor for primary care, which we foster from infancy to adulthood with integrated healthcare data. With the AI doctor, we can solve the inefficiency problem. In stage three, after we have proven that the autonomous mobile clinic network can truly solve the target clinical use cases, we shall open up the platform for all medical verticals, thus enabling universal healthcare through this whole new system.
Shaoshan Liu is the founder and CEO of PerceptIn, an autonomous machine computing company. His research interests include computer architecture, deep learning infrastructure, robotics, and autonomous driving. Liu has published more than 100 research papers and holds over 150 U.S. and international patents. He is also the lead author of the textbooks Creating Autonomous Vehicle Systems, Engineering Autonomous Vehicles and Robots: The DragonFly Modular-based Approach, and Robotic Computing on FPGAs. His career goal is to improve humanity through technological advancements.
Liu was an ACM Distinguished Speaker from 2019-2022, where he gave talks on topics including 3D scene reconstruction, robotic localization, and designing systems to keep autonomous vehicles safe.