MSc Research: Robot Learning (in progress)
Direct responsibilities:
Research design, Policy learning experiments, Empirical evaluation.
Hands-on contributions:
Imitation learning baselines, Sim-to-real evaluation protocols, Open-source training tooling, Empirical evidence on data efficiency.
About the project:
Research at UCL exploring how autonomous robots can learn generalisable manipulation and navigation skills from limited human demonstration. The work investigates how foundation models, imitation learning, and sim-to-real transfer can reduce the data and engineering effort needed to deploy robots in unstructured real-world environments.
Abstract
Robot learning has shifted from hand-engineered controllers toward data-driven policies that map perception directly to action. Recent advances in foundation models, imitation learning, and large-scale simulation have made it plausible to train general-purpose robotic skills, but deploying these policies on real hardware in unstructured environments remains brittle, data-hungry, and difficult to evaluate.
Recent work in robot learning has begun to address this through three complementary directions: pretraining visuomotor policies on diverse demonstration data so that downstream tasks need fewer examples, using simulation and domain randomisation to expose policies to long-tailed conditions before real-world deployment, and structuring training pipelines so that human teleoperation, autonomous rollouts, and corrective demonstrations all contribute to a single improving policy. Each direction shows promising results in isolation, but the practical question of how to combine them into a reliable training loop for a specific robot and task is still largely open.
The core problem this research addresses is that there is no clear methodology for deciding, in a given robotics deployment, how to allocate effort between collecting demonstrations, training in simulation, and refining policies on real hardware in order to reach a target success rate with minimum cost.
By grounding policy learning in measurable data efficiency and sim-to-real transfer, this research aims to support faster and more trustworthy deployment of autonomous robots into real-world settings. The outcomes are relevant to robot learning, embodied AI, and human–robot interaction, and are intended to contribute to venues such as CoRL, RSS, and ICRA.