Deploying intelligence and autonomy in constrained edge platforms

Where existing solutions for autonomy are unsuitable for your edge platform, from drones to ground robotics, a novel approach to sensing, fusion, intelligence and system engineering might be what you need.

Automation is programming a system and constraining the environment so that repeating the same actions time and time again achieves your goal. Autonomy is creating a system that reliably achieves this goal in a changing environment. Autonomous systems need to:

Tightly integrating sensors, algorithms and actuators allows efficient performance. When you understand what you really need to see and what you can ignore, you can design a perception system that is custom tuned to your application. Where more information is needed, adding different types of sensor and combining their outputs is often the way forward. Sensor fusion techniques allow you to build systems with greater robustness and an understanding of the environment far richer than a single sensor could provide. Closed loop feedback from your actuators allows you to get the most out of this understanding, moving with precision and efficiency.

Modern AI, whilst reducing the barrier to entry, provides a complex design space for Sense and Think. Foundation models have astonishing capability in extracting information from data but have equally weighty computational requirements. Do you need to operate a datacentre full of GPUs so your farming robots can avoid workers in the field? If you do, what does it cost to transmit your data from the field back to the datacentre and the answers back to the field? At what point does the latency of connectivity make it slower to get answers than a simpler computer running on your robot? These are complex techno-economic questions, and the right answer changes all the time as communications, compute, and models evolve at a rapid pace.

Safety is also critical. Many of the most exciting and impactful applications of autonomy are also the applications where we have always relied on the judgement of a human to ensure safety: aviation, mobility, construction and mining. When we hand over the controls to an autonomous system, do we require that system to be safer than the human it replaces, or just ‘as safe’? And critically, how do we prove this safety to the satisfaction of a regulator? Classical machine learning is often ‘explainable’, but when a neural network is involved, how can you be sure it will always give the answer you need, and what definition of always do you care about? Formal proof methods might be some of the answer, as may be digital twins, model-based development, and effective guard-rails.

When fine tuning a foundation model or training a bespoke network you can get great initial results, but maintaining this performance as the technology is deployed and scales is far more challenging. A great AI system is designed from the ground up for scalability. How could you efficiently collect, process and store data? How do you ensure that this data is high enough quality and sufficiently representative of the environment your system will be used in? What happens when the colours of an object fade and change, leaves fall off the trees and road markings disappear? Simulations, synthetic data, and unsupervised learning can all be used help to reduce this burden, but the architecture needs to include provisions for this from the start. With expectations of AI soaring, starting with the right approach has never been more important

Size, weight, power and cost. These are the eternal trade-off for a systems engineer. Somehow you must balance the demands to carry more, travel further, and cost less with the need to embed the sensors and compute that allow you to achieve your goal. When pushing to the limit, a whole-system perspective is needed. Will adding a higher resolution sensor mean a simpler algorithm can be used, saving compute power, or does this add weight and cost without improving the outcomes? Does fusing the output of two simple sensors allow a cheaper and more robust system overall? When do you have to freeze the interfaces to allow the subsystems to get on with the job of designing and delivering? As always, best is the enemy of ‘good enough’, and it takes years of experience to know how to navigate the pitfalls.

From aviation to mobility and broader industrial technology, we provide our expertise and experience in sensor fusion, algorithms, artificial intelligence and system engineering to help our clients embrace and develop technologies for autonomy that will keep them ahead of the curve and their competition.

Contact us to learn more about how TTP can accelerate your development.