News · missionSTUDIO research
The swarm now has a reflex.
Introducing missionREFLEX: the fast-decision direction built from our swarmREFLEX research. Agentic intelligence can frame the mission. Physical AI still needs bounded decisions at the pace of the floor.
Explore the approach ↓Enabling The Sovereign Frontier
One mission. Three timescales.
A frontier model can help interpret a situation and examine alternatives. A mixed fleet also needs to assign work, avoid a crossing and respond to changing positions without waiting for a long model call. Our research separates those jobs.
Know the world.
Relate places, assets, people and observations so a decision has the context of the operation.
Coordinate the mission.
Propose fleet assignments and routes, with agentic supervision where an override can be reviewed.
Respond in time.
Keep narrow, time-critical decisions close to the fleet, with deterministic constraints and specialised models evaluated against simple baselines.
What we have built and tested
Logic for hard limits. Models for incomplete context.
The swarmCORTEX simulator includes a swarmREFLEX arm with hand-written, reciprocal velocity-avoidance logic. It is a simulated control rule, not a learned model. Exact fleet assignment provides another deterministic reference for testing narrow decisions.
Separately, a small joint assignment model was trained against the exact solver and evaluated on held-out simulation seeds. At a 10-second tag age, it agreed with the exact teacher on 64.8% of per-agent decisions, compared with 51.1% for an always-nearest heuristic on the degraded view.
Those numbers measure decision agreement in an offline simulation evaluation. They do not measure field safety, throughput or a model operating inside the live simulator. missionREFLEX is the product direction for taking this work toward site-specific integration and validation.
The next step
Build the capability. Prove the boundary.
missionSTUDIO is being developed as the place to model an operation, test a decision path and review evidence before a field trial. A local response layer belongs in a real mission only after its inputs, timing, constraints and human authority are verified for that site.