First video of Opteran natural intelligence in action
Today is an important day for the team at Opteran and one for celebration. Since the start of the year, we have been working incredibly hard to deliver against our technology milestones and we’re delighted to share the first video of our technology in action. In the video you can see Hopper, our robot dog (named after Grace Hopper, a pioneer of computer programming) moving around a course using components of Opteran Natural Intelligence – not a trained deep learning neural net. Our small development kit (housing an FPGA) sat on top of the robot dog guides Hopper, using Opteran See to provide 360 degrees of stabilized vision, and Opteran Sense to sense objects and avoid collisions.
What’s wrong with autonomy today?
Why do we believe this is an exciting development? First, we should look at what’s wrong with today’s state-of-the-art artificial intelligence (AI). The most common limitations with models, like deep learning neural networks, have been widely discussed by the likes of Gary Marcus. They require extensive training to recognize objects and are subject to bias in training, as well as struggle to generalize. This means ever increasing amounts of compute power and expanding data center footprints. Furthermore, integrators often have to throw away investments in solutions to keep up with whichever neural network appears to be “flavor of the month.” This lack of consistency makes it hard to achieve scalability and means robustness is ever harder to achieve with each change. This makes the cost-benefit analysis for widespread AI adoption difficult to justify, as well as raising concerns about its environmental impact.
Nature, on the other hand, has been solving these challenges for 600 million years and more. Where existing AI systems struggle with a mismatch between what engineering thinks it needs to solve and optimize, nature has found simpler but more robust solutions to the same problems. In reality nature has outperformed today’s deep learning and Mike Davies, director of Intel’s neuromorphic computing lab, said at the end of last year: “Brains really are unrivalled computing devices…Quantitatively, nature outperforms computers three-to-one on all dimensions…Today’s computer architectures are not optimized for that kind of problem…The brain in nature has been optimized over millions of years…deep learning can only capture a small fraction of the behavior of a natural brain.”
So why does nature do it better?
Firstly, nature does movement and decision making better as all the elements have co-evolved, so brains, bodies and behavior work efficiently together. Today’s AI technology relies on a wide array of software, sensors and hardware, which are costly and inefficient when compared to nature’s solution.
Secondly, nature tends to find optimal solutions to get the task completed against all of the variables of a particular scenario. For example, nature extracts the visual cues necessary to complete a task, pre-processing data very fast in early visual systems and only feeds the right information at the right time. Deep learning models must process every single piece of information before making a decision.
Thirdly, nature uses multiple sensing modalities to extract static and temporal cues which affect decisions. All of this is done subconsciously, i.e. autonomously, and at orders of magnitude lower size, weight and power. For instance, in navigation and collision avoidance tasks insects have evolved coupled sensing and processing algorithms to extract multiple, redundant cues that when combined correctly allow them to achieve incredible navigational feats.
Beyond Level 5 Autonomy
The limitations with first and second generation AI technology means that systems are only just beginning to reach Level 4, which means true autonomy and understanding is beyond the reach of the technology. This means companies are having to experiment with new techniques to improve autonomy, but this repetitive cycle makes robustness harder to achieve. We believe that Opteran Natural Intelligence is the answer to these failings, as (like Nature) it is constantly evolving to build more robustness which will see us move beyond level 5 autonomy to true full autonomy.
To qualify what we mean by “full autonomy,” we have re-evaluated how to define it, and are specifying it in terms of general abilities, rather than level of automation of a particular task (e.g. driving). As such we see levels 1-5 as mapping to our See, Sense, Direct capabilities and then beyond that point we envisage level 6 as machines observing and understanding the world around them, a first example of which is Opteran Decide and is only one aspect of level 6. Then, level 7 will see machines interacting with the world around them, which is when we believe machines will achieve full autonomy.
This is possible, because Natural Intelligence mimics nature’s co-design of sensors, processing and actions for a holistic solution using filtered sensing cues to deliver the right algorithms for the right task. Consequently, we have an autonomy solution that is operationally efficient and ensures mission continuity thanks to the robustness and redundancy built into our platform. In the near future, it will also enable greater scalability as its certified and verified algorithms will give confidence that outputs will be accurate and require far less data processing.
Additionally, the unique end-to-end Opteran Toolchain enables rapid continuous innovation in going from natural behavior to real-world application in silicon without the need for costly redesign, costly data curation and training. Solutions can be simulated in weeks and realized in real-world products in a fraction of the time of other technologies.
Why is Natural Intelligence such a breakthrough.
Today we see many examples showcasing sensor and technology ‘overfit’ in the field of autonomous robots and machines. Engineers are constrained by existing AI technologies, which leads to complex and computationally expensive solutions to problems that nature solves far more efficiently. By mimicking nature and its far more efficient, robust approach to autonomy, Opteran Natural Intelligence is delivering general purpose solutions for a wide range of autonomous machines and addressable markets.
UAVs are a particularly good example as low in-flight SWaPs (size, weight and power) are critical to achieving increased operational efficiency. Lower power and weight means increased operational time and increased operational time can make the difference between needing one instead of two UAVs to complete a task in the available time.
One use-case we analysed was an indoor warehouse inventory management drone. It had to be small to navigate autonomously through aisles with optimal coverage of all shelves / storage areas it had not yet seen, whilst being safe and not hitting shelves or the contents of shelves. However, despite the need for compactness using existing deep learning technologies it had to employ 5x stereo depth cameras and a heavy-weight GPU for the navigation and task processing – roughly 400g for sensors alone.
We are also finding many others making drones, such as for logistics, which are using LiDAR based flight systems weighing from 1-1.5Kg. By contrast we find that our Natural Intelligence approach is capable of achieving more robust autonomy with a single system with orders of magnitude less weight, size and power. Consequently, we believe we are the only solution that will be able to carry out full omnidirectional collision avoidance and visual only navigation on sub 250g drones.
In the future, given our in-depth research into nature and insect brain function, our approach will also make it easier to integrate self-awareness and decision making. Adding such functionality will be feasible, as insect brains are not overly complex so it will not significantly increase power consumption or the processing requirements of the algorithms. Consequently, we believe this will open up a new frontier for autonomous robotics!
