Research
We work on the algorithmic and engineering problems that arise when many robots have to act together.
Our work spans four overlapping themes — multi-agent coordination, learned communication, vision-based navigation, and the lab-scale hardware platforms that make experiments possible.
Themes
Learned communication
End-to-end learning of what to say, when, and to whom — under bandwidth, latency, and reliability constraints.
Multi-agent coordination
Decentralised decision-making and emergent cooperation in robot teams under partial observability and constrained communication.
Swarm robotics
Hardware platforms and field experiments for multi-drone research at lab scale.
Vision-based navigation
Lightweight visual perception and control for resource-constrained aerial and ground robots.
Papers
2026
- Apr 7PAPER ICRA 2026Communication-constrained collective intelligence for indoor drone swarms
Bourached, Smith, Prorok
We study how teams of small quadrotors can maintain collective task performance when communication links become sparse, lossy, or actively jammed. Our approach learns a compact, task-relevant message protocol jointly with the control policy, and we evaluate it in indoor swarm experiments on the Argus platform.
2025
- Nov 2025PAPER RA-LOn-board visual ego-motion for sub-100g quadrotors
Smith, Prorok
A lightweight visual-inertial odometry stack designed to run within the power and compute budget of a sub-100g indoor quadrotor.
- Oct 2025PAPER NeurIPS 2025Heterogeneous multi-robot reinforcement learning
Bettini, Shankar, Prorok
We study cooperative multi-robot tasks where the team is composed of agents with structurally different sensors, actuators, and compute budgets. We introduce a centralised-training, decentralised-execution scheme that explicitly conditions on agent type, and show that it scales to teams of twelve heterogeneous robots with strictly better performance than role-blind baselines.