Unmanned Systems Technology 006 | ECA Inspector Mk2 USV | Antenna systems | Northwest UAV NW-44 | Unmanned ground vehicles | Navigation systems | Lunar X challenge

82 February/March 2016 | Unmanned Systems Technology PS | Navigation software A s we reported in the previous issue of UST , simultaneous localisation and mapping (SLAM) navigation technology is ideal for applications where an autonomous vehicle is in an environment where a GPS-based navigation system cannot communicate with satellites, and/ or where the system’s accuracy isn’t sufficient for the vehicle to operate safely (writes Stewart Mitchell). However, the technology has an issue with recognising the vehicle’s location when it tries to navigate in surroundings that are changing in real time, such as urban pedestrianised areas, because it uses the location of objects as reference points to map the area around it. This is where a technology called SINS (self-improving navigation system) could be used. It is not a new technology – its development began more than 20 years ago – but it may start appearing in autonomous vehicles in the near future. SINS uses a combination of high-speed robotic control software and multiple-layer machine learning algorithms that allow an autonomous system to adapt to changing environments. Its algorithm structure is far more complex than SLAM, as at the core of its layers is a constructive conceptual change mechanism that progressively constructs symbolic mathematical representations of the environment in the vehicle’s field of view. These are called ‘experience identifiers’, and they are logged by the system’s onboard memory, and every new experience is cross- referenced with each of the previous experiences to identify any similarities between each one. This means that instead of just logging the objects around it, it logs its whole ‘experience’, which includes the shapes and sizes of the objects around it as well as their vectors, to identify permanent and non-permanent objects. Built into these identifiers is a series of representational reaction values that denote the vehicle’s physical space options that it can move into. These values allow the vehicle to dynamically select an appropriate robotic control response to a situation, while feeding real- time reaction data back to its memory and adapting its chosen response to the new environment as it moves forward. As the vehicle repeats experiences, the system recognises patterns and begins to categorise them along with the vehicle’s action-reaction log, which speeds up the reaction time of the system as it narrows the ultimate number of action possibilities down to one. This categorisation also means subtle changes in the environment will not occupy large amounts of memory by creating a new memory with very small changes; instead the system will recognise experience overlap and add or subtract the differences between ‘then’ and ‘now’. For autonomous systems to operate safely using SINS, a huge amount of computing power and memory is needed, as well as a ‘rules’ filter that will be cross-referenced with every decision made by the system to ensure the vehicle doesn’t go rogue. As SINS requires very powerful hardware, it will probably only ever be used by large autonomous vehicles, but wouldn’t it be interesting to see self- improving autonomous vehicles in the near future? Well, watch out for SINS. Now, here’s a thing “ ” Instead of just logging the objects around it, SINS logs the whole ‘experience’ of the vehicle’s environment