MANO
A 6-DOF arm on a mecanum base that takes instructions in plain language — Raspberry Pi 5, on-camera inference, and a spatial-reasoning model in the loop.
Why build it
I have spent years making software that never touches anything. I wanted a project where the feedback loop ends in a physical object either being picked up or being knocked over, because there is no arguing with that.
MANO — Mobile Autonomous Navigation Operator — is a mobile manipulator built out of hobby-grade parts on a budget under four hundred dollars.
The architecture
Split the real-time from the reasoning. An ESP32 drives the servos through a PCA9685 and owns everything with a deadline: PWM timing, the ultrasonic and time-of-flight sensors, the IMU loop. The Raspberry Pi 5 owns everything without one: vision, planning, talking to the model. They speak over serial with a small command protocol. A vision model stalling should never mean a servo holding a stalled position under load.
Inference on the camera. The IMX500 runs its network on the sensor and emits detections rather than frames. On a 4GB Pi that is the difference between a robot and a space heater.
Language in, poses out. Gemini Robotics-ER takes the scene and an instruction and returns spatial reasoning — what is where, what to grasp, in what order. The Pi turns that into joint angles. The model is a planner, not a controller; nothing in the closed loop waits on a network call.
Current state
Chassis drives. Arm moves through its full range under inverse kinematics. Grasping works on objects it has seen before, at known heights, on a flat surface — which is to say, in the easiest possible world. Everything after this is the hard part.
Where it’s going
I’m documenting the build publicly, and I’m interested in whether a hardware-plus-software subscription is a viable shape for a small team. That question is unresolved and I would rather say so than pretend I have a business model.