Variational Inference as a Framework for Intelligent Robots

Introduction
Intelligent robotic systems present several challenges to successfully develop and deploy in the real world. A variety of technical sub-problems need to solved and integrated elegantly. For instance, a self-driving car needs to perform perception (process the incoming data from its sensors), control (maintain the desired velocity and the appropriate steering angle), planning (decide which route to take) and many more. Experts in different areas have developed a
variety of methods for each of the sub-challenges. However, first, the integration of these different methods is a non-trivial task and has become an area of study on its own. Second, in all of these areas, state-of-the-art robots can't reach
human-level performance yet. In this project, it's proposed to use methods from the computational neuroscience literature to provide a framework to solve a large set of sub-challenge in robotics. Specifically, the free-energy principle and it's extension to Active Inference proposed by the renowned neuroscientist, Karl Friston. Friston, a professor at University College London, proposed the free-energy principle less than a decade ago. Since then it has acquired significant prominence in computational and systems neuroscience as a general theory of brain and behaviour. This principle claims that in order for an intelligent agent to persist in a time-varying environment, it must minimize `surprise' (the atypicality of an event). This is done by minimizing an upper bound called `Free Energy'.
Subsequent research has shown the validity of this method in explaining a large set of capabilities of biological agents. This includes perception, action selections, planning under uncertainty and learning. Additionally, it has shown to outperform methods such as reinforcement learning for benchmark problems.

Aims and objectives
The aims of this project can be divided into two sections. First, further development of the Active inference framework. This includes:
- Establishing a formal connection between Active Inference and current state-of-the-art methods in robotics.
- Developing rich approximation methods that are computationally feasible for robots.
- Formal proofs of the performance (why does this method work and can we rely on it to always work).
Second, apply this framework to robotic systems. In this project, the focus will be on wheeled mobile robots. This includes:
- Present algorithms for control, perceptions, planning and learning for mobile robots.
- Benchmark this method against state-of-the-art methods.
- Analysis of the integration problem.

Novelty and potential impact
Despite this great potential, this framework has not been used in robotics. There are only a handful of papers published on Active Inference applied to a robot. All of those include positive but preliminary results with toy examples. Successful implementation of this method in robotics will boost the state-of-the-art in multiple research areas and will provide a general framework for a large set of problems. This is unprecedented in the field of robotics. Additionally, this framework provides the best-known model from neuroscience to explain the human brain. This could be a leap in artificial intelligence towards human-level intelligence.

EPSRC strategy alignment
This research is relevant to the EPSRC strategies. It ts best under 'Information and communication technologies (ICT)' and specifically Artificial Intelligence, Robotics and modern statistics. All these areas are indicated to grow.