Research

My research focuses on enabling autonomous decision-making and control in legged robotic systems operating in complex, uncertain, and dynamic environments. I am particularly interested in how robots can learn robust and adaptive locomotion strategies that allow them to perceive environmental changes, select appropriate behaviours, and execute stable movement without continuous human supervision.

Rather than treating locomotion purely as a tracking problem, I study it as an adaptive control problem in which behaviour must remain reliable across variations in terrain, robot dynamics, sensing noise, and task demands.

Research Interests

Within this broader vision, my work concentrates on learning-based control architectures for legged robots, with emphasis on the following challenges:

1. High-dimensional embodied control: Legged robots are complex multi-DOF systems with strongly coupled dynamics and hybrid contact interactions, making stable and efficient control difficult.
2. Robustness in uncertain environments: Real-world deployment requires locomotion policies that remain effective under terrain changes, disturbances, model mismatch, and partial observability.
3. Adaptation and behavioural flexibility: Beyond robustness, I am interested in controllers that can express and switch between meaningful locomotion behaviours depending on context, task, or environmental conditions.
4. Structured internal representations: A central question in my work is how control policies represent gait dynamics, environmental context, and task-relevant information internally, and how such representations can support improved adaptability and generalisation.

Current Research Areas

1. Deep Reinforcement Learning for Legged Locomotion: Developing modular and structured control architectures for legged robots, with an emphasis on robustness, behavioural diversity, and generalisation across tasks and environments.
2. Simulation-to-Real Transfer: Investigating methods to transfer learned locomotion policies from simulation to hardware while accounting for uncertainty, unmodelled dynamics, and discrepancies between training and deployment conditions.
3. Representation Learning for Control: Studying how low-dimensional latent representations can be learned from high-dimensional sensory and proprioceptive inputs, and how these representations relate to behaviour, adaptation, and policy performance.
4. Robust and Adaptive Locomotion: Analysing how training design, task variation, and privileged or latent information can improve the ability of legged robots to remain stable, recover, and adapt their motion across changing conditions.