Shahab Kaynama

I have moved to Clearpath Robotics since Novermber 2014. Page last modified on 29/08/2015.

Contact:   [lastname] or

I live and breathe controls and robotics. My work lies in the intersection of control systems, computer science, and mathematical optimization. Between 2012-14 I was a postdoctoral researcher at UC Berkeley under Prof. Claire Tomlin. I received my Ph.D. in 2012 in Electrical and Computer Engineering from the University of British Columbia where I worked under Profs. Meeko Oishi, Ian Mitchell, and Guy Dumont. I have an M.Sc. (2006) in Advanced Control and Systems Engineering from the University of Manchester. Outside of work I enjoy playing chess, reading books, playing the guitar, and exercising. My favourite sport is tennis.

The Safe Learning Project

I get asked a lot of questions about this project, so here is a quick snippet of what it is all about.

Most machine learning algorithms in robotics either completely lack formal guarantees of safety, or else impose stringent conditions on the system. For example, they might assume that we can always safely rely on the observed data; that our knowledge of the environment or the system improves monotonically, and that our uncertainty constantly decreases as we collect more data. The problem is that in the real world one can never fully trust observed data.

This is where our safe-learning algorithm shines: It provides a systematic method to preserve safety for generic nonlinear dynamics even in the existence of uncertainty or our lack of sufficient observation. As a result, we can run our favourite machine learning algorithm to improve performance of the system and have the peace of mind that the system could never go unstable or jeopardize safety (both of itself and its surrounding environment). For more information please see this paper.

Stuff I do is so stuff like this never happens (again!):

The STARMAC platform, Hybrid Systems Lab, UC Berkeley, Feb. 2014.
Creative director: The talented J. F.-Fisac.

Safety Analysis and Synthesis in Robotics

Computing the set of states in which existence of a safety-preserving control law is guaranteed is extremely desirable when it comes to ensuring safe operation of robotic systems (particularly if those systems are to operate among humans; for example in the case of self-driving cars or materials handling robots on the factory floor). Yet, providing guarantees of safety is a challenging task in high dimensions. The main problem is due to Bellman's "curse of dimensionality": the complexity of computing the safe set increases exponentially with the dimension of the state.

My research in this area focused on finding scalable techniques for a class of robotic systems such that 1) we can efficiently compute a correct (conservative) approximation of the safe set, and 2) synthesize the corresponding safety-preserving control laws in a manner that is permissive. This allows us to "wrap" the safety algorithm around any existing control infrastructure and immediately guarantee a safe operation of the system.

Selected Publications (updated sporadically)

Doctoral Thesis:

International Journal Articles:

Refereed International Conference Proceedings:

Copyright (c) 2008-14