CerebrumFlow is platform-agnostic. It ships as the cognition layer for the machines that do real physical work — from warehouse fleets to field robots to autonomous vehicles. The loop is the same; the platform below it changes.
Arms and mobile bases working together on dexterous, unscripted tasks. Whole-body control resolves arm and base as one system, so a pick that's slightly out of reach produces a coordinated base shift and reach — not a two-step shuffle. Force/torque feedback means contact tasks comply instead of jamming, and re-grasps happen on the fly when a part shifts.
Autonomous mobile robots navigating dynamic floors — re-planning around people, pallets, and each other without a central pause. Fused lidar-camera SLAM keeps the map current as stock moves, and fleet re-planning hands off zones between robots so a dropped pallet or a stalled unit doesn't ripple into a line-wide stop.
Quadrupeds and crawlers in unstructured outdoor environments — energy, agriculture, construction — where the map is never fixed and the ground never flat. Fused visual-inertial odometry holds position where GPS is denied or degraded, and uncertainty-aware perception lets the robot flag what it can't confidently read instead of reporting a wrong answer.
Perception-to-action reasoning that has to be right the first time, at speed. The planner keeps a rolling horizon of candidates and prunes anything that violates a collision bound, re-planning at 40 Hz so a closing gap is handled as it forms. Deterministic scheduling means latency is a guarantee, not a best effort.
CerebrumFlow isn't a from-scratch rewrite of your robot. It sits on top of the middleware and motor interfaces you already have.
Publishes and subscribes as a native ROS 2 node over DDS, so it composes with existing perception, drivers, and tooling without an adapter layer.
Talks to EtherCAT and CAN-based drives for joint-level torque and position, keeping the control loop close to the motors.
Consumes cameras, lidar, IMU, and joint encoders through standard drivers — calibration and time-sync handled in the runtime.
Import your robot's URDF or MJCF, and the dynamics model and kinematic constraints are derived automatically.