Robot eyes are power hungry. What if we gave them tools inspired by the human brain?
- Written by Adam D Hines, Research Fellow, Centre for Robotics, Queensland University of Technology
Robots are increasingly becoming a part of our lives – from warehouse automation to robotic vacuum cleaners. And just like humans, robots need to know where they are to reliably navigate from A to B.
How far, and for how long, a robot can navigate depends on how much power it consumes over time. Robot navigation systems are especially energy hungry.
But what if power consumption was no longer a concern?
Our research on “brain-inspired” computing, published today in Science Robotics, could make navigational robots of the future more energy efficient than previously imagined.
This could potentially extend and expand what’s possible for battery-powered systems working in challenging environments such as disaster zones, underwater, and even in space.
How do robots ‘see’ the world?
The battery going flat on your smartphone is usually just a minor inconvenience. For a robot, running out of power can mean the difference between life and death – including for the people it might be helping.
Robots such as search and rescue drones, underwater robots monitoring the Great Barrier Reef, and space rovers all need to navigate while running on limited power supplies.
Many of these robots can’t rely on GPS for navigation. They keep track of where they are using a process called visual place recognition. Visual place recognition lets a robot estimate where it’s located in the world using just what it “sees” through its camera.
But this method uses a lot of energy. Robotic vision systems alone can use up to a third of the energy from a typical lithium ion battery found onboard a robot.
This is because modern robotic vision, including visual place recognition, typically relies on power-hungry machine learning models, similar to the ones used in AI like ChatGPT.
By comparison, our brains require just enough power to turn on a light bulb, while allowing us to see things and navigate the world with remarkable precision.
Robotics engineers often look to biology for inspiration. In our new study, we turned to the human brain to help us create a new, energy-efficient visual place recognition system.
Mimicking the brain
Our system uses a brain-inspired technology called neuromorphic computing. As the name suggests, neuromorphic computers take principles from neuroscience to design computer chips and software that can learn and process information like human brains do.
An important feature of neuromorphic computers is that they are highly energy-efficient. A regular computer can use up to 100 times more power than a neuromorphic chip.
Neuromorphic computing is not limited to just computer chips, however. It can be paired with bio-inspired cameras that capture the world more like the human eye does. These are called dynamic vision sensors, and they work like motion detectors for each pixel. They only “wake up” and send information when something changes in the scene, rather than constantly streaming data like a regular camera.
These bio-inspired cameras are also highly energy efficient, using less than 1% of the power of normal cameras.
So if brain-inspired computers and bio-inspired cameras are so wonderful, why aren’t robots using them everywhere? Well, there are a range of challenges to overcome, which was the focus of our recent research.
A new kind of LENS
The unique properties of a dynamic vision sensor are, ironically, a limiting factor in many visual place recognition systems.
Standard visual place recognition models are built on the foundation of static images, like the ones taken by your smartphone. Since a neuromorphic sensor doesn’t produce static images but senses the world in a constantly changing way, we need a brain-inspired computer to process what it “sees”.
Our research overcomes this challenge by combining neuromorphic chips and sensors for robots that use visual place recognition. We call this system Locational Encoding with Neuromorphic Systems, or LENS for short.
LENS uses the continuous information stream from a dynamic vision sensor directly on a neuromorphic chip. The system uses a machine learning method known as spiking neural networks. These process information like human brains do.
By combining all these neuromorphic components, we reduced the power needed for visual place recognition by over 90%. Since nearly a third of the energy needed for a robot is vision related, this is a significant reduction.
To achieve this, we used an off-the-shelf product called SynSense Speck, which combines a neuromorphic chip and a dynamic vision sensor all in one compact package.
The entire system only required 180 kilobytes of memory to map an area of Brisbane eight kilometres in length. That’s a tiny fraction of what would be needed in a standard visual place recognition system.
Hexapod robots have six legs and can walk on different surfaces both indoors and outdoors.A robot in the wild
For testing, we placed our LENS system on a hexapod robot. Hexapods are multi-terrain robots that can navigate both indoors and outdoors.
In our tests, the LENS performed as well as a typical visual place recognition system, but used much less energy.
Our work comes at a time when AI development is trending towards creating bigger, more power-hungry solutions for improved performance. The energy needed to train and use systems like OpenAI’s ChatGPT is notoriously demanding, with concerns that modern AI represents unsustainable growth in energy demands.
For robots that need to navigate, developing more compact, energy-efficient AI using neuromorphic computing could be key for being able to go farther and for longer periods of time. There are still challenges to solve, but we are closer to making it a reality.
Authors: Adam D Hines, Research Fellow, Centre for Robotics, Queensland University of Technology
