Plymouth research · 2008–2014
I was an early pioneer applying GPU computing to cognitive robotics. I used graphics processors to make larger robot-learning experiments practical, working with the iCub humanoid and building the Aquila research toolkit.
My 2014 Plymouth thesis brought this work together. Its abstract describes it as the first investigation of GPU computing’s impact on cognitive robotics.
Question and contribution
Learning an action means learning a sequence, not recognising one photograph. Larger neural models could connect more sensory information, movements and words, but their training cost constrained what researchers could try. A GPU runs many calculations in parallel; the question was how to use that capacity for embodied learning.
During my Plymouth PhD, I developed CPU and CUDA GPU versions of neural-network code for iCub control and training. I worked across the learning algorithms, experiments and Aquila tooling. By July 2012, we had implemented multi-GPU training and tested it on four-GPU systems.
Multiple time-scales recurrent neural networks (MTRNNs) learn sequences at different rates, relating slower patterns to faster movements. Back-propagation through time (BPTT) trains these networks across a sequence. Aquila brought CPU/GPU modules together with graphical controls and visualisation, so researchers could run and inspect experiments.
Aquila 2.0 architecture paper, 2013Aquila also supported related work in the lab: Morse, Belpaeme, Cangelosi and Floccia identify it as the software used for their 2011 study of U-shaped learning curves.
Reported result
In NVIDIA’s historical CUDA Spotlight interview, I reported roughly a 50-fold speed increase for the iCub training algorithms when comparing my GPU and CPU implementations. Faster training made larger experiments practical.
This was a comparison of those implementations, with optimisation still incomplete. The later four-GPU tests were a separate experiment.
Read my original comparison
Collaboration
I worked with supervisors Angelo Cangelosi, Tony Belpaeme and Davide Marocco, and co-authored the Aquila 2.0 paper with Anthony Morse and Angelo Cangelosi. The iCub platform and the wider action-language research gave us a robot and a scientific setting in which to test the ideas.
The linked iCub film records preliminary tests combining object and voice recognition with neural-network training and execution in Aquila 2.0.
Separate research strand · 2009–2010
Alongside the GPU work, I took part in the Plymouth–ESA Evolution in Robotic Islands study and co-authored its report. We asked whether populations of candidate neural controllers could improve by exchanging solutions.
We paired Plymouth’s Mars Rover Simulator with ESA’s PaGMO island-model optimiser. Cangelosi, Marocco and Barry Bentley contributed on the Plymouth side; Christos Ampatzis and Dario Izzo co-authored the final report. The simulator and 2010 report preserve the experiment.
This was a separate evolutionary-robotics strand: the ESA framework used parallel CPU threads, while my CUDA work accelerated neural-network learning.
The 2009 rover-controller paper is also cited in Stefan Kral’s 2011 TU Berlin diploma thesis, preserved in DLR’s electronic library.
NVIDIA and teaching
I presented my CUDA cognitive-robotics work at NVIDIA’s Santa Clara headquarters on 29 September 2011, then wrote about the visit on 1 October.
Read the original presentation accountI returned to Silicon Valley for a research internship at NVIDIA in summer 2012. My preserved intern badge and career record sit alongside the public research from that period.
Alongside my research, I held part-time lecturing and demonstrating contracts at the University of Plymouth. I prepared and taught CUDA classes and worked on the GPU computing lab. This teaching continued until February 2014, separately from the NVIDIA internship and conference speaking.
Public artifacts
My thesis and supplemental iCub, Aquila and rover videos.
Read the thesis and supplementsThe preserved toolkit history and original releases, with CPU/GPU modules and experiment controls.
Browse the GitHub archiveThe GTC research poster, alongside the separately dated multi-GPU implementation note.
