Shawn Singh

SteerSuite is a set of tools, code, and test cases for developing steering behaviors. If you have any interest in steering behaviors, then please take a look! If you are planning to prototype or develop your own steering AI, you will definitely want to use SteerSuite to save yourself from implementing a lot of infrastructure.

SteerSuite web page: www.magix.ucla.edu/steersuite.

SteerBench is a benchmarking tool for steering behaviors. It is now part of SteerSuite.

We created a steering algorithm that uses egocentric fields - that is, fields centered around each agent. The egocentric property allows us to avoid resolution and storage problems associated with global field-based steering methods. It also allows us to perform space-time planning, because time to reach any given point can be approximated by distance from the origin (the agent's location). With this approach, agents show natural, predictive behaviors in a broad variety of situations.

"Egocentric Affordance Fields in Pedestrian Steering," M. Kapadia, S. Singh, W. Hewlett, P. Faloutsos. In Proc. Interactive 3D Graphics and Games, I3D, 2009.   paper   movie

One of the main problems with research in steering behaviors is that each paper demonstrates its results on its own focused set of scenarios, and usually only performs subjective evaluation. SteerBench consists of a diverse, challenging suite of test cases, metrics, and a method of objectively scoring steering behaviors without knowledge of an algorithm's internal workings. SteerBench software will be made available soon.

"SteerBench: A Benchmark Suite for Evaluating Steering Behaviors," S. Singh, M. Kapadia, P. Faloutsos, G. Reinman. In Computer Animation and Virtual Worlds, invited to a special issue on gaming, to appear 2009.   paper   movie

SIMD (vector) instructions, a form of fine-grain data parallelism, were one of the primary contributions that made real-time ray tracing possible (e.g., Ingo Wald's PhD thesis). In this paper, we slightly change the photon mapping algorithm to make it possible to use SIMD instructions for photon mapping as well. The resulting framework was able to achieve near-interactive performance for low-resolution images.

"SIMD Packet Techniques for Photon Mapping," S. Singh, P. Faloutsos. In Proc. Symposium on Interactive Ray Tracing, RT07 , 2007.   paper   movie

Photon mapping is a popular algorithm for rendering images with high quality global illumination. In this paper, we propose a conceptual architecture for photon mapping and analyze the throughput and bandwidth requirements. Results indicate that photon mapping could run at interactive rates on today's hardware, and with future generations of processors, could be real-time. While this paper originally proposed to accelerate photon mapping in hardware, the analysis applies to a software photon mapping architecture as well, which is more practical to consider.

"The Photon Pipeline Revisisted," S. Singh, P. Faloutsos. In The Visual Computer, vol. 23, no. 7, pp 479-492, 2007.   paper   movie

To do any sort of automatic computation, abstract computations are represented by physical processes. For the past fifty years, the most common bridge between physical phenomena and abstract computation has been to model the transistor (or many other devices) as a switch. Networks of switches can do amazing abstract computations, such as adding or multiplying binary numbers, and much, much more. In my opinion, this switch abstraction is not necessarily the most efficient way to compute. I am interested in exploring many other ways to use physics for computation. One particular approach I believe in, is to make physics to perform constraint satisfaction. At the nanometer scale, this sort of paradigm could not only save power and work faster than networks of transistors, but also could be capable of significantly more abstract computation per unit area. Examples of novel computations using constraint satisfaction at the nanoscale are Quantum-dot Cellular Arrays (QCA), DNA computing, ground-state computing, and in some ways, quantum computing. I hope to search for useful ways to compute using constraint satisfaction by approaching from the theoretical side, exploring the intersection of nanoscale physics, discrete dynamical systems, theory of computing, and boundary-value PDEs.

Of all 3-d display technologies, only holograms can simultaneously achieve (1) high resolution parallax, (2) support for multiple viewers, and (3) no need for special viewing glasses or tracking devices. However, there is a huge number of challenges to solve, in display technology and software algorithms, before holographic displays are truly practical. On the software and architecture side, very little work has explored the possibility of real-time algorithms for rendering to a holographic fringe pattern, and it will be a challenge to manage the sheer bandwidth of terapixels per second. There are even more challenges on the display technology side, which I am still learning.

Given a machine with various sensory inputs and various ways to interact with its environment, is it possible to create a program that gives the machine human-like learning, awareness, and cognition? This question is very controversial and interesting, and I have some vague ideas how such a program might work.