Evidence over assurances
A good demo is not proof. A policy has to beat scripted, random and idle baselines on fresh random seeds, with controls, before we call it progress. When it does not, we say so.
About
We build the intelligence that lets robots work as a team.
Iresine is a robotics software company focused on one hard problem: coordinating many robots at once. We train that coordination with multi-agent reinforcement learning in simulation, test it against simple baselines and build it to run on each robot’s own computer.
Our mission
Coordination should be learned, not scripted.
Getting one robot to move well is hard but well understood. Getting dozens of them to share a job is a different problem. Each robot sees only part of the world, messages arrive late or not at all, and the right move for one robot depends on what every other robot is doing.
Most fleets today run on hand-written rules or a single central controller. Both work well in controlled settings, and both get harder to maintain as teams grow and the world stops behaving like the test site.
That is why Iresine exists. We train each robot’s policy (the part that decides what to do next) with multi-agent reinforcement learning in simulation, so the team works out how to cooperate through practice. At run time there is no central brain to wait on.
Three ways to run a fleet
-
Hand-written rules Brittle
Work for a handful of robots in a test area, then break as the fleet grows and conditions change.
-
One central controller Fragile
Every decision waits on one computer and its radio link. It is a bottleneck and a single point of failure.
-
Learned, decentralized coordination Our approach
The team practices together in simulation. In the field, each robot decides for itself from its own sensors and local messages.
What we focus on
Multi-agent reinforcement learning, built for swarms.
Reinforcement learning trains behavior by practice: a robot tries actions, gets a score for the result and gradually improves.
Multi-agent reinforcement learning (MARL) does this for a whole team at once, so each robot learns behavior that helps the group, not just itself. We work on it through four connected pillars.
-
Simulate
Swarm-scale simulation with physics and sensor models, many environments running in parallel and randomized conditions.
-
Train
Centralized training, decentralized execution: a shared critic sees the whole team while learning, but each robot acts on its own sensors and messages.
-
Coordinate
Learned communication under bandwidth limits, decisions with partial information, and teamwork that holds across team sizes and when robots drop out.
-
Deploy
Sim-to-real transfer into small onboard networks, inside a safety layer of constraints, geofences and fallback behaviors.
Evaluate: every policy is compared against scripted, random and idle baselines on fresh random seeds before anyone trusts it. How we evaluate
Explore the technologyPrinciples
How we work
The commitments behind every environment we build and every policy we train.
Decentralized by design
Robots should keep working when the network or a teammate does not. We train for local sensing and local messages, so no single controller can stop the whole team.
Safety is part of the system
Learned behavior runs inside a safety layer of constraints, geofences and fallback behaviors. We design that layer alongside the policy, not after it.
Built with the people who run the robots
Operators know the rules, failure modes and edge cases no simulator starts with. That is why a project with us starts by defining the mission with them, and why field experience feeds back into training.
Who it’s for
Built for teams with real multi-robot problems.
Teams come to multi-agent learning from different starting points. Here is what each typically needs.
Robotics companies
Building multi-robot products
Typically needs
- Coordination that scales as fleets grow
- Policies small enough to run on each robot
- A clear path from simulation to field tests
Research labs
Studying multi-agent learning
Typically needs
- Swarm-scale simulation environments
- MARL training and evaluation tooling
- Honest baselines and repeatable seeds
Public-safety teams
Emergency response and search
Typically needs
- Coverage of large areas that change fast
- Behavior that holds up when radio links are poor
- Clear safety limits and fallback behaviors
Industrial operators
Running fleets day to day
Typically needs
- Fewer hand-tuned rules to maintain
- Steady performance when robots drop out
- Software that works with the robots they already run
Where we’re headed
Early, focused and looking for partners.
Iresine is an early-stage company. We are building the core platform for simulating, training and evaluating robot teams, and we test it in our own research environment, starting with wildfire response: teams of agents learn to locate, track and contain spreading fire in simulation.
Next, we want to work with partners who have real multi-robot problems, so the platform is shaped by missions that matter in the field rather than by benchmarks alone.
Now
- Building the simulation, training and evaluation platform
- A first research environment: wildfire response, in simulation
- Testing every policy against baselines before we trust it
Next
- Pilot projects with teams that build or run multi-robot systems
- Research collaborations on multi-agent learning
- Field data flowing back into simulation
Building a robot fleet that needs to think as one?
Tell us about your robots, your environment and the behavior you need. We will tell you honestly whether multi-agent learning can help, and how.