The University of Plymouth thesis record describes this as the first investigation of GPU computing’s impact on cognitive robotics. CUDA made it practical to scale recurrent neural systems for embodied AI experiments with the iCub humanoid robot.
The evidence spans the iCub and multiple time-scales recurrent neural networks, Aquila’s heterogeneous CPU/GPU architecture, multi-GPU back-propagation through time, a separate Plymouth–ESA evolutionary-rover collaboration, and NVIDIA records from 2011 onward.
Martin Peniak with the iCub humanoid robotGPU acceleration
Research line
Compute changed which questions were practical.
Embodied learning needed larger experiments. The public record shows the neural models, implementation work, open tooling, measured acceleration, and the separate evolutionary-rover research that formed the wider path.
iCub + MTRNN
Action, language, and recurrent dynamics
Large multiple time-scales recurrent neural networks connected perception, language, and full-body iCub action sequences.
Multi-GPU BPTT
Training across several GPUs
Aquila implemented multi-GPU training of MTRNNs with back-propagation through time; a contemporaneous 2012 record documents tests across four GPUs.
Plymouth + ESA
Evolutionary rover controllers
A separate island-model research strand evolved neural controllers for planetary-rover navigation in collaboration with ESA’s Advanced Concepts Team.
Aquila + CUDA
Reusable heterogeneous experiments
The open architecture made CPU/GPU cognitive-robotics modules easier to run, distribute, teach, inspect, and accelerate.
Research work
What the robotics work made visible.
The work connected robot sensing, body control, recurrent neural networks, language grounding, simulation, and GPU acceleration. The important part was not one algorithm. It was the operating loop that made experiments possible.
CUDA acceleration changed the scale and iteration speed of iCub neural models.Aquila joined reusable CPU/GPU modules with multi-GPU MTRNN training.A 2012 archive record documents four-GPU BPTT tests and the start of GPU-based fitness-evaluation work for scalable genetic algorithms.The ESA strand evolved rover neurocontrollers through an island-model framework; it was distinct from the CUDA work.
Plymouth lab
The iCub work was a research community.
This photograph places the work inside the Plymouth iCub lab with supervisor Angelo Cangelosi, colleagues, student and co-author Barry Bentley, and the robot at the centre. The thesis, Aquila tools, action-language videos, rover work, and GPU acceleration came from that living research environment.
The work reached NVIDIA at the 2012 GPU-AI inflection point.
An invited 2011 presentation at NVIDIA headquarters set out where parallel GPU computing could help cognitive robotics. A research internship at NVIDIA in Silicon Valley followed in summer 2012, the same year the AlexNet result made GPU deep learning globally visible. The documented record supports that technical contribution and timing without claiming that one researcher caused NVIDIA’s wider AI strategy.
This photograph shows the public presentation stage for the same line of work: Aquila, CUDA acceleration, iCub cognitive robotics, and the practical tooling needed to make larger embodied-learning experiments possible.
Martin’s contemporaneous technical note records an implemented multi-GPU BPTT version for MTRNN training, tests on four-GPU systems, and early work on scalable genetic algorithms using GPU-based fitness evaluation.
When an experiment takes too long, the research question quietly shrinks. GPU acceleration and reusable tools widened the space of possible robotics experiments. That concern with practical constraints later reappears in edge cameras, synthetic scenes, and spatial calibration.
Thesis
NVIDIA HQ
NVIDIA Santa Clara
NVIDIA GTC 2014
Aquila
ESA rover work
NVIDIA keynote
Robotics
Rover simulation
Active vision
