Magnus Egerstedt: From Coordination to Collaboration in Multi-Robot Systems
Harvard SEAS Computer Science Colloquium Series, 15 October 2026
Table of Contents
Event
| Field | Value |
|---|---|
| Talk | From Coordination to Collaboration in Multi-Robot Systems: Lessons from Ecology |
| Speaker | Magnus Egerstedt, Executive Vice Chancellor and Provost, University of North Carolina at Chapel Hill |
| Series | Computer Science Colloquium Series, Harvard SEAS |
| Date | EDT |
| Venue | Science and Engineering Complex (SEC) LL2.224, 150 Western Avenue, Allston MA 02134 |
| Format | Colloquium, open to the public |
| URL | events.seas.harvard.edu |
The argument
The standard recipe for a robot team is decomposition. A team-level goal is broken into pieces any one robot can handle, such as holding a formation or sweeping a region, and the robots run those pieces in coordination. Each robot does what it could have done alone, only in the right place at the right time.
The talk asks what happens when the team is heterogeneous and the pieces are not interchangeable. A ground robot and an aerial robot together may reach states neither can reach alone, so the team gains capabilities none of its members has. Ecology has a name for this kind of arrangement: mutualism, where two species each get something from the other that they could not get by themselves.
The formal move is to treat each robot's capabilities as a safe set, encoded by a barrier function, and to model collaboration as composing those functions so that one robot's safe set expands with the help of another, at some cost to the helper's. The setting is long-duration autonomy, where staying alive and operational matters more than finishing a task optimally. The abstract names three settings: environmental monitoring, safe learning, and the Robotarium, a remote-access multi-robot lab that has run almost without pause for more than eight years.
Background: Egerstedt's work on the topic
Control barrier functions
A control barrier function certifies that a set of states is forward invariant: start inside it and the closed-loop system stays inside. Ames, Xu, Grizzle and Tabuada (2017) (Ames et al. 2017) made this practical by posing the safety condition as a linear constraint on the control input, so a quadratic program finds the input closest to a nominal controller that still satisfies it. Ames, Coogan, Egerstedt, Notomista, Sreenath and Tabuada (2019) (Ames et al. 2019) is the survey of the theory and its applications, robotics among them, and the usual entry point.
The pattern matters for this talk because it separates the task from the constraint. The nominal controller does the job; the barrier function only intervenes when the job would leave the safe set.
Safety barrier certificates for teams
Wang, Ames and Egerstedt (2017) (Wang, Ames, and Egerstedt 2017) applied the same quadratic-program filter to collision avoidance across a team, first centrally and then in a decentralized form where each robot only has to stay safe relative to its neighbours. The experiments used nominal controllers deliberately designed to make the robots collide. Glotfelter, Buckley and Egerstedt (2019) (Glotfelter, Buckley, and Egerstedt 2019) extended nonsmooth barrier functions, which already supported Boolean composition of constraints, to constraints that change over time and jump, such as an obstacle that appears when sensed. Composition of barrier functions is the mechanism the abstract leans on.
The Robotarium
The Robotarium at Georgia Tech is a multi-robot testbed anyone can use remotely: users write a controller against a simulator, in Python or MATLAB, and submit it to run on the physical robots. Pickem et al. (2017) (Pickem et al. 2017) describe the design. The hard requirement is that uploaded code must not be able to damage the hardware, which calls for safety routines that intervene as little as possible while still carrying guarantees: barrier certificates again. Wilson et al. (2020) (Wilson et al. 2020) report what years of open access taught them.
A lab where strangers' code drives the robots is a long-duration autonomy problem in miniature. The safety layer has to hold for any controller, not just a well-behaved one.
Robot ecology and long-duration autonomy
Egerstedt, Pauli, Notomista and Hutchinson (2018) (Egerstedt et al. 2018) set out the program: for robots that must operate for months, design by constraints rather than by objectives. Egerstedt's book Robot Ecology (2021) (Egerstedt 2021) develops it at length. Its central borrowing from ecology is that an organism's behaviour is shaped by the constraints of its habitat, and it replaces optimal task policies with survivability constraints that keep the robot alive.
Persistification is the plainest example. Notomista and Egerstedt (2021) (Notomista and Egerstedt 2021) make "the battery is never empty" a barrier-function constraint, so exploration or surveillance tasks run far past one battery's life, with experiments on the Robotarium.
Collaboration in heterogeneous teams
Nguyen, Jabbari and Egerstedt (2024) (Nguyen, Jabbari, and Egerstedt 2024) look like the most direct source for the talk. They study pairwise collaboration between robots with different dynamics, encoding each robot's safe operating region with a control barrier function. A helping robot can let a helped robot enter regions it could not traverse alone. The paper defines a safely reachable set (where a robot can get safely, with or without help), gives conditions under which a helper and a helped robot can pair up, and structures the arrangements as hybrid automata, with experiments on mobile robots. "Why, when and how should robots collaborate?" is the paper's own framing.
Safe learning
Two papers connect the barrier-function machinery to learning. Emam et al. (2022) (Emam et al. 2022) handle disturbances such as hardware wear or terrain by modelling them as a union of convex hulls, which lets a learned disturbance model, a Gaussian process for instance, feed robust barrier functions. Emam et al. (2025) (Emam et al. 2025) put a differentiable robust barrier-function layer inside model-based reinforcement learning, so exploration never leaves the safe set and the reward can be learned without reference to the constraints.
Speaker
Magnus Egerstedt is Executive Vice Chancellor and Provost and Harry Watson Distinguished Professor at the University of North Carolina at Chapel Hill. Per the SEAS listing, he was previously Stacey Nicholas Dean of the Samueli School of Engineering at UC Irvine, and before that at Georgia Tech, where he chaired Electrical and Computer Engineering and directed the Institute for Robotics and Intelligent Machines. He studied at KTH Royal Institute of Technology (M.S. in Engineering Physics, Ph.D. in Applied Mathematics) and at Stockholm University (B.A. in Philosophy), and was a postdoc at Harvard. His research is in control theory and robotics, especially the coordination of multi-robot systems. He is a Fellow of IEEE and IFAC, a member of the Royal Swedish Academy of Engineering Sciences, and a past president of the IEEE Control Systems Society.
The biography is taken from the listing and not checked independently.
Questions to bring
- Mutualism in ecology is often conditional and can tip into parasitism. Does the composition framework model a collaboration that stops paying off for the helper, and who decides to break it?
- Composing barrier functions can produce an empty feasible set. What does a team do when two robots' safe sets can't both be kept, and is that a scheduling problem or a control problem?
- Pairwise collaboration is analysed in the 2024 paper. Do the results carry to three or more robots, or does the hybrid automaton grow too fast?
- After eight-plus years of the Robotarium, which failures got past the safety layer, and were they model errors (the dynamics were wrong) or specification errors (the safe set was wrong)?
- Survivability over optimality is also the stance of a production system under an SRE. Does the robot ecology framing say anything about software agents that run for weeks?
Notes
Notes pending.