How to avoid using resource-intensive cloud technologies to AI-powered avatar movements control?

There is no doubt that large neural networks, including LLM, are highly promising for generating animations of avatar movements. But their real time usage - to improve the quality of avatar movement control in virtual reality - remains problematic. This is especially a bottleneck for mass users entering VR simulations by using such low-budget devices as a virtual helmet combined with a mobile phone. The PAO-XR startup implemented compact ML-databases - instead of large-scale neural networks - individually trained for each user and localized on the user's device. This allows to provide AI-powered avatar movements control in real time from cost-effective user's devices and without cloud technologies.