Technology

A cognition loop, not a pipeline.

Traditional robotics stacks are a chain of hand-tuned modules — perception feeds prediction feeds control — where an error anywhere becomes an error everywhere. CerebrumFlow replaces that chain with a single learned loop that keeps sensing, reasoning, and motion coherent, end to end.

01 / PERCEIVE

Perception with uncertainty

We fuse cameras, lidar, IMU, and joint encoders into a single probabilistic scene. Every estimate carries calibrated uncertainty — a per-estimate covariance that says how wrong it might be — so the system knows what it doesn't know, and passes that knowledge downstream instead of dropping it at the module boundary.

This is the part most stacks get subtly wrong. An occluded corner is not empty space. A dust-obscured lidar return is not a wall. Treating "no signal" as "nothing there" is how robots drive confidently into the thing they couldn't see.

cameras · lidar · IMU · proprioception → sceneestimate + uncertainty band
  • sensor fusionCameras, lidar, IMU, and joint encoders fused into one scene, without assuming every sensor agrees — disagreement is evidence, not noise.
  • fused odometryVisual-inertial odometry blended with wheel and lidar odometry for ego-motion that stays tight through tunnels, dust, and GPS-denied indoor spaces.
  • relocalizationLearned place recognition for loop closure and recovery after a kidnapping or a battery swap — the robot re-anchors to the map without a cold restart.
  • scene understandingSemantic and geometric segmentation that labels things and stuff — a pallet, a gap, a person mid-stride — not just raw pixels.
  • occlusion handlingOcclusion and dropout treated as first-class measurements: when a sensor goes blind, its contribution shrinks toward the prior instead of vanishing.
  • calibrated uncertaintyHeteroscedastic uncertainty — the model outputs how wrong it might be, per estimate — carried into planning rather than discarded at the interface.
02 / PLAN

Constrained reasoning in milliseconds

Planning here is model-predictive control over a short receding horizon. We roll out a few hundred milliseconds of candidate trajectories against a learned dynamics model, score them against cost and safety, and commit only the first action before the world moves again.

Constraints aren't penalties tuned by hand — they're hard. Joint limits, velocity and acceleration caps, and the non-holonomic constraints of a wheeled base are enforced as feasibility filters, so an unsafe candidate never reaches the controller.

goal + constraints → action candidatesfeasible + safe selection
  • MPC horizonA receding horizon of a few hundred milliseconds, rolled out against a learned dynamics model — only the first action is committed.
  • hard constraintsKinematic and dynamic limits — joint ranges, velocity and acceleration caps, non-holonomic base constraints — enforced as feasibility, not tuned penalties.
  • safety filterCandidates that violate a keep-out zone, a collision bound, or a torque limit are pruned before they ever reach the controller.
  • continuous replanningReplanning every cycle — if a person steps into the corridor between two rollouts, the next plan already accounts for them.
  • learned cost shapingCost shaping learned from demonstrations, so natural motion emerges without scripting every affordance by hand.
  • task-level reasoningReasoning over what to do next, not just how to get there — chaining subgoals when a direct path is blocked.
03 / ACT

Closed-loop actuation

Intent becomes motion through a feedback loop that measures and corrects in real time. Whole-body control resolves arm and base together — a reach that needs the base to shift produces one coordinated motion, not a two-step shuffle. Errors are corrected within the cycle they occur, not accumulated until failure.

Contact is where most motion stacks come apart. For insertion, assembly, and grasping, we run impedance control with force/torque feedback, so the robot complies instead of jamming when a part is a millimeter off.

intent → motion → feedback40 Hzerror corrected, not accumulated
  • whole-body controlArm and base motion resolved together, so a reach that needs the base produces one coordinated trajectory.
  • impedance controlForce/torque feedback for insertion and contact tasks — the robot complies instead of jamming when a part is off by a millimeter.
  • joint-level controlTorque and position control with real-time interpolation at the loop rate, corrected within the cycle an error appears.
  • per-joint limitsTorque, velocity, and thermal limits enforced in the controller with graceful derating instead of a hard fault.
  • motion primitivesLearned motion primitives that retarget to a different arm, payload, or base without a rewrite of the task intent.
  • safe stop & recoveryA safety-rated stop if an external force exceeds threshold, recovering without a full system reset.
The runtime

Edge-first, deterministic, on a clock.

The loop doesn't touch the cloud. Perception, planning, and actuation run on embedded compute under a real-time scheduler with a fixed latency budget — so "fast" is a guarantee, not a hope.

loop rate
40 Hz

Perceive–plan–act completes 40 times a second, the rate real motion needs.

latency budget
24 ms

Perception 11 ms, planning 8 ms, actuation 5 ms — a fixed, enforced budget.

compute
edge SoCs

Jetson-class and automotive SoCs. No cloud round-trip on the control path.

scheduling
deterministic

Real-time scheduler with watchdog. A stage that overruns degrades, it doesn't stall.

degradation
graceful

If a stage misses its budget, the loop drops to slower replanning and keeps the safety filter running.

integration
ROS 2 / DDS

ROS 2, DDS, and EtherCAT for motor-level I/O — drops into existing robot stacks.