Research

Robots are dependable when the world is arranged for them, and unreliable the moment it is not. The gap opens wherever contact matters — fitting a part that was never measured, handing an object to a person, working with material that changes shape under the grasp. Classical control carries the guarantees but needs a model; learning brings the flexibility but offers no promises.

My work sits between the two, drawing on control theory, differential geometry and machine learning to build manipulation that is adaptable enough for an unstructured world and principled enough to be trusted in it. The recurring theme is structure: the geometry of the space a robot moves through, the physics of the contact it makes, and the guarantees a system should carry before it works alongside people. The thrusts below are the fields this runs through; trajectory optimization and the identification of robot dynamic parameters run underneath most of them.

Research thrusts

A geodesic across a curved surface, with the tangent space and a tangent vector at the start point tangent space goal
Thrust 1

Geometric and Riemannian robot learning

Orientations, stiffness matrices and robot configurations are not flat vectors. They carry curvature and constraints. My work formulates learning and control directly on these manifolds. Generalization then respects the structure of the space instead of working around it.

A learned trajectory spiralling into an equilibrium along Lyapunov level sets while avoiding a hatched unsafe set unsafe set V(x) level sets
Thrust 2

Safe and certifiable learning-based control

A controller that usually works is not the same as one that provably will not fail. My work establishes stability and safety guarantees. These guarantees still hold when the components producing the behaviour were learned from data rather than designed.

Several demonstrated trajectories inside a variance band converging into one bold reproduced motion shown reproduced
Thrust 3

Learning from demonstration and physical interaction

A robot has to come to have a skill in the first place. It can be shown by a person, or discovered through its own exploration and contact with the world. The skill must then survive the move to new objects, new goals and different bodies.

A tool held by a spring and damper pressing a fingertip into a surface that deforms under the contact force F stiffness ยท damping
Thrust 4

Compliant and contact-rich manipulation

Most useful tasks are defined by contact rather than free motion. My work regulates force and impedance for assembly, insertion, tool use and the handling of deformable objects. The aim is to keep the robot stable and safe while it touches something, or someone.

A continuum arm whose curvature grows along its length, curling around an object object base
Thrust 5

Soft and continuum robotics

In these robots the compliance lives in the body rather than in the controller. My work models, simulates and controls continuum structures. Their effectively infinite degrees of freedom put them out of reach of classical rigid-body methods.

A fan of possible future trajectories with a widening cone of uncertainty around the chosen one now chosen
Thrust 6

Decision-making under uncertainty

Acting well means acting without an exact model of the world. My work uses free-energy and active-inference formulations. They treat exploration and robustness as two faces of one problem, not as competing objectives to be traded off.

Muscle activity decoded into intent that drives a prosthetic hand, with sensory feedback returning to the arm muscle decode intent prosthesis feedback
Thrust 7

Neural interfaces and assistive robotics

My work reads human intent from biological signals, meaning muscle and brain activity. It then closes the loop through prosthetic and wearable devices that respond to that intent. The sensory feedback they give makes them feel like part of the body.

Selected projects

FREE-MPC — Free-Energy-Guided Safe Manipulation Under Uncertainty
NYU Research Catalyst Grant — internal, awarded.
DR-FREE — learning-based robust control via free energy
Free-energy formulation of robust robot control, with a public code release.