Research / Featured research

Learning-Based Control

We combine machine learning with physics-based control so robots can move with agility and adapt to new conditions, while their behaviour stays interpretable.

In the media: Caltech News

IROS · Oral presentation · 2025

Quadrotor morpho-transition: learning vs model-based control strategies

I. Mandralis, R. M. Murray, M. Gharib

Landing by transformation drives a quadrotor to the edge of actuator saturation. On hardware, an end-to-end reinforcement learning policy lands with more agility, but only once motor dynamics and sensing delays are part of its training world. Model-predictive control transfers out of the box, yet recovers less gracefully when actuators fail unexpectedly.

Paper arXiv

Block diagram of the training loop: policy, motor dynamics, body dynamics and simulator — **Training on what the hardware feels.** The learned policy π(s) acts through first-order models of the motors and of the morphing body, so that thrust, moment and tilt commands reach the simulated robot with the same lags as the real one.

Time-lapse of ATMO landing on its wheels under a learned policy, with position, velocity and attitude plots — **A learned landing.** Under the RL policy, ATMO descends, folds its rotors and touches down on its wheels at a 65° tilt. Roll and pitch stay within a few hundredths of a radian until the moment of impact.

Nature Communications · 2021

Learning efficient navigation in vortical flow fields

P. Gunnarson, I. Mandralis, G. Novati, P. Koumoutsakos, J. O. Dabiri

Reaching a target through unsteady currents with only local information. Reinforcement learning finds navigation policies for a fixed-speed swimmer in a cylinder wake that use the flow rather than fight it. What the swimmer senses matters: sensing velocity outperformed bio-inspired vorticity sensing, reaching the target nearly every time and approaching the travel time of optimal trajectories.

Paper Caltech News

A swimmer's learned path through the vortex wake behind a cylinder, with a close-up of the local flow field — **Riding the wake.** A swimmer steered by a learned policy crosses the unsteady wake of a cylinder to reach its target (A). Up close (B), its path weaves between counter-rotating vortices, taking the currents that carry it forward.

More research Ground-Aerial Morphing Robotics · Fluid-Body Interaction and Flow Control · Vision-Based Navigation and Manipulation