howdy!

picture of my face
I'm Eric Undersander and I'm a research engineer at Facebook AI Research. We're applying AI to robotics tasks like navigation and mobile manipulation. We're building an open-source platform to accelerate this work, including a fast 3D simulator.

My background includes motion planning for Cruise's self-driving car, performance optimization for deep neural networks at Baidu's Silicon Valley AI Lab, and engine development in the video game industry.



View my LinkedIn profile

block sparsity

Recurrent Neural Networks (RNNs) are widely used in speech recognition, machine translation, and language modeling. However, the current trend in deep learning towards larger datasets and larger models has led to increased storage requirements and inference latency.

In this research at Baidu SVAIL, my colleague Sharan Narang and I explored block sparsity, in which unimportant blocks of network connections are pruned. Block-sparse RNN models reduce storage and compute requirements while still achieving good accuracy and high compute utilization on deployment hardware.

demo video

A quick montage of some of my early solo video game projects:

intelligent camera

I wrote a technical article about the third-person camera we built for Kingdoms of Amalur: Reckoning, with emphasis on my obstacle avoidance solution. The Sticky Collision Beam was published in the September 2011 issue of Game Developer Magazine.

blog posts

Here's the takeaway of this whole post: For camera collision zoom, don't cast a ray. Don't cast a sphere. Cast the near face of the view frustum. Read at Gamasutra.com
I’ve used Havok Physics at several game studios over the years and I thought I’d share some little-known but hopefully useful tips about this middleware. Read at Gamasutra.com

contact me



View my LinkedIn profile