Aim

This PhD project focuses on developing a novel hybrid control framework that combines Model Predictive Control (MPC) and Behavioural Cloning (BC) to enable robust, real-time locomanipulation for navigating and surveillance of complex structures using legged robots. The project aims to train legged robots with motor skills that enable them to perform autonomous surveillance. Specifically, we focus on enabling dynamic motions through MPC, and the versatile behaviour needed to open doors or clear path tasks through BC in quadruped and humanoid robots equipped with manipulators.

Objectives

1. Design contact-implicit stochastic MPC frameworks that can efficiently compute policy gradients and handle uncertainty in contact events;
2. Develop a novel diffusion-based learning framework for MPC controllers that can clone dynamic behaviours such as opening doors and path clearance during surveillance operations;
3. Integrate BL techniques with MPC controllers to execute dynamic surveillance operations autonomously; and
4. Apply this integrated framework to real-time surveillance on steel and cluttered structures with legged robots.

Description

Model Predictive Control (MPC) has demonstrated remarkable capabilities in enabling agile robotic behaviors—most notably, dynamic maneuvers such as backflips in Boston Dynamics’ Atlas robot. However, conventional MPC methods remain constrained by local optima, limiting their ability to plan complex motion and contact sequences, particularly in cluttered or uncertain environments. These limitations are especially evident in loco-manipulation tasks, where both mobility and interaction with the environment are required.
 

Moreover, existing whole-body MPC frameworks are largely deterministic, which makes them ill-suited for real-world uncertainties, especially in contact-rich scenarios. On the other hand, behavioural cloning (BC) via diffusion policies has recently shown impressive success in learning the diversity of manipulation behaviors but struggles to scale to more whole-body behaviours where balance and dynamics are critical. Fundamentally, diffusion policies capture and learn the complex distributions present in human behaviours by learning the de-noising process from collected data.

This PhD research will explore a hybrid approach, combining the structure and real-time feasibility of MPC with the flexibility and autonomous capabilities of BC, to enable robust and versatile surveillance on legged robots. Concretely, this project aims to enable legged robots to move around complex environments, requiring them to open doors and remove debris autonomously.

The project builds on advances in robot motor intelligence, differential contact simulation, and model predictive control developed internally in the RoMI lab. It will also advance our current research efforts in Neural Conditioning Probability (NCP) for behavioural cloning.

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. 

Further information on English language requirements for EU/Overseas applicants.

Home fee rate and stipend are available for this position.