Fig. 2: Key differences in informational requirements for training and deploying natural and artificial solutions to autonomy.

Transformer-based approaches to autonomy rely on internet-scale datasets for training, and process input from a suite of high resolution sensors, such as 4k cameras and LiDar, in order to provide a limited behavioural repertoire in comparison to biological autonomous agents. In contrast, 600 m years of evolution on a planetary scale, with complex physics, has encoded blueprints to build autonomous brains into the genome of a massive variety of animal species. These brains process much sparser input from specialized sensor suites, captured using active perception and behaviour, to generate a hugely rich variety of adaptive behaviours. Image sources: NASA (Earth), insectbraindb.org (honeybee brain, reproduced under https://creativecommons.org/licenses/by/4.0/).