Steering AI out of harm’s way

16 October 2019

For Dr Yinlam Chow (Croucher Scholarship 2014), having an interest as a child in robots and all they could do turned out to be a future career in the making because today he is a research scientist at Google AI, working on artificial intelligence (AI)-enhanced algorithms and applications that enable robots to navigate their way safely around their environment.

When Chow was growing up in Hong Kong, he found himself drawn to find out more about machines such as robots and aircrafts. This led him to enrol in the Department of Mechanical Engineering at the University of Hong Kong (HKU), where he had the opportunity to join ABU Robocon competitions that require team members with knowledge across disciplines – mechanical and electronic circuit design, and programming.

Through HKU’s student exchange programme, he spent a year studying aeronautics at Embry-Riddle Aeronautical University, Florida, and after graduation in 2009 decided to return to the US to pursue a master’s degree in aeronautics and astronautics at Purdue University.

While at Purdue, Chow developed research interests in nonlinear control systems and optimization theories, in particular control algorithms in areas such as aircraft stabilisation and auto-pilot navigation.

This research experience also led on to a new interest in machine learning, and for his doctoral studies Chow sought to study at Stanford University’s Institute of Computational and Mathematical Engineering (ICME). There, supported by his Croucher Scholarship, he delved into sequential decision-making, risk-sensitive optimal control, and reinforcement learning, exploring applications related to robot motion planning, fleet management, and renewable energy systems. Chow also gained industry experience at Xerox PARC, Adobe Research, and IBM Smarter Transportation Center.

Keen to continue on the applied side of research, in 2016 Chow joined the founding team of Osaro, Inc, a start-up focused on developing machine learning systems for industrial robots. Although he went there as a research engineer/scientist, in Osaro’s early days there were multiple roles for Chow to manage, ranging from developing algorithms for Osaro Pick – a proprietary machine learning robot pick-and-place system – to leading reinforcement learning-based pilot projects.

After 18 months, an opening arose for Chow to join DeepMind, an Alphabet company focusing on AI research in reinforcement learning. While at DeepMind, he was primarily focused on projects related to the safety and robustness of sequential decision-making.

In his “Lyapunov-based Safe Policy Optimization” project, for example, Chow and his collaborators studied continuous action reinforcement-learning problems where it is crucial that the agent only interacts with the environment through safe policies, i.e. policies that do not take the agent to undesirable situations. He went on to develop novel reinforcement-learning algorithms that were more data-efficient and effective than other leading methods.

Chow also collaborated with Google Brain Robotics to deploy these algorithms on the indoor navigation platform of a personal assistant robot. The task set for the robot was to learn to reach a destination as efficiently as possible while limiting total impacts with objects to avoid harm to its surrounding environment, including any people nearby, and to itself. “In motion avoidance, if there is a route from point A to B, we need to make sure there is a safety threshold of very close to 100 percent,” Chow pointed out. A video of the robot navigating its way through an office can be viewed here.

Today, Chow is working at Google Research to tackle optimal control problems in robot navigation (visual-servoing) and dialog management, in which the input data is high-dimensional and the underlying dynamics model is not known in advance. In the future, he hopes to enhance the human-agent interactions of Google’s recommender systems, such as Assistant and YouTube, to facilitate real-time conversations. To do so, he is seeking to leverage artificial intelligence to help these systems better understand what people are talking about and to make more personalised recommendations.

This could take the form of starting conversations and interacting with people as well as recommending locations and advertisements that may interest them. “I want future ‘assistants’ to have a more dynamic role in people’s lives,” he said.



Dr Yinlam Chow is a research scientist at Google Research. Prior to Google, he worked at DeepMind from 2017-19 and at Osaro, Inc from 2016-17. He received his BEng degree in Mechanical Engineering with first class honours from the University of Hong Kong in 2009 and his MS degree in Aeronautics and Astronautics at Purdue University in 2011. He completed his PhD at the Institute of Computational and Mathematical Engineering (ICME), Stanford University. Chow received a Croucher Scholarship in 2014.


To view Dr Yinlam Chow’s Croucher profile, please click here