Nengo and Low-Power AI Hardware for Robust, Embedded Neurorobotics

Area of Science:

• Robotics
• Neuroscience
• Computer Science

Background:

• Developing robust, embedded neurorobotic systems faces challenges in environmental interfacing, sensory processing, control signal generation, and hardware compilation.
• Existing tools often require specialized knowledge, hindering rapid development and deployment.

Purpose of the Study:

• To demonstrate how the Nengo neural modeling and simulation libraries facilitate the rapid development of robotic perception and action neural networks.
• To showcase Nengo's capability for simulating these networks on neuromorphic hardware using familiar tools like Keras and Python.
• To address the primary challenges in building embedded neurorobotic systems.

Main Methods:

• Utilizing NengoInterfaces for environment and sensor interaction.
• Employing NengoDL with Keras and TensorFlow APIs for neural network development.
• Implementing the Neural Engineering Framework for white-box function and circuit design.
• Leveraging Nengo's backend libraries (e.g., NengoLoihi) for hardware compilation.

Main Results:

• Successfully developed and simulated spiking neural networks for a simulated rover's target tracking and control using Nengo.
• Integrated Nengo for neural adaptive control augmentation in a real-world robotic arm task, improving performance.
• Demonstrated the compilation and execution of Nengo models on CPUs, GPUs, and Intel's Loihi neuromorphic chip.

Conclusions:

• Nengo provides a unified framework to overcome key challenges in neurorobotic system development.
• The libraries enable efficient development and deployment of neural networks for robotic applications across diverse hardware platforms.
• Nengo empowers researchers to build and run their own neurorobotic systems with accessible tools and detailed implementation guidance.