PhD in Biomedical Engineering - Brain inspired robotic planning

About the Project

One of the most challenging aspects of robotic navigation is to achieve reliable and robust decision-making while navigating through unknown environments. For example, it is crucial for a rescue robot to be able to navigate an unfamiliar building to the rescue site and back without becoming stuck. However, both the traditional autonomous-hoover-SLAM and the current the state of the art called "end-to-end training" need the robot to visit a place many times to create a map. However, how can one navigate without a map and we human do that with ease. We don't need a map when entering a coffee shop we've never seen before. End to end training tries to overcome this by creating a 3D model of the entire world and then train a network on it. Obviously, that has it's problems, too:

• End-to-end training is very expensive and power hungry to train, due to the amount of data and computation required, and also power-hungry to deploy, due to the size of the neural networks used.
• It is trained offline before being deployed on the robot and might not be able to adapt to unseen conditions at runtime.
• The learned offline model is a black box and thus is not explainable, including its world model or the decision-making strategy used.
• Learning from scratch involves learning even basic physics concepts such as collisions each training run, while every gaming engine has a fundamental understanding of the world.

Towards a solution inspired by neuroscience and psychology:

Studies in the Psychology and Neuroscience of humans and animals have revealed that they use mental simulations to rehearse their options when making decisions in unfamiliar or uncertain conditions. An element which may facilitate or enable these simulations is an observed innate understanding of principles such as collisions, solidity and causality, which is called "Core Knowledge" in psychology. In this mental simulation, this may enable the animal to, for example, predict a collision with an upcoming object and make a turn to avoid it, rather than only knowing to do so next time. We propose to reproduce this core knowledge-informed mental simulation in an autonomous robot vehicle at runtime, as an alternative to costly offline trial-and-error simulations.

A proof of concept (Lafratta et al 2025, https://direct.mit.edu/neco/article-abstract/37/7/1288/131054/Closed-Loop-Multistep-Planning) has been successfully deployed on a real robot (https://github.com/berndporr/bratmobile). However, this PhD will have the aim to arrive at a system which is able to master full tasks in large space such as factory floors and is able to learn from its experience. The proposed research objectives impact world-leading research by defining a new approach to robot navigation and decision making. The proposed technique would have a low implementation cost with a small carbon footprint but offers high reliability, explainability and safety. It therefore has potential to have a wide range of applications and drastically lower the cost of task automation.

Thus, this PhD aims to:

• Perform mental simulations in real time on the robot using a validated physics engine to explore the robot's current world model to ensure sound and explainable multi-step reasoning.
• Represent behaviours as adaptive, attentional, transient feedback controllers; this representation reduces the computational load drastically by focusing the robot's attention only on target locations or crash sites.
• Chain low level feedback controllers to create plans and to store them as on-board cognitive maps which are explainable and can then be retrieved in similar situations.
• Create and maintain a real-time open-source C++ library for real robots with a modular, user-friendly and customisable architecture, with the goal to allow beneficiaries to adapt it to their needs and hit the ground running with their robot projects.

Reference: Giulia Lafratta, Bernd Porr, Christopher Chandler, and Alice Miller. Closed-loop multistep planning. Neural Computation, 37(7):1288–1319, 06 2025. ISSN 0899-7667. doi: 10.1162/neco a 01761. URL https://doi.org/10.1162/neco_a_01761.