A robot that learns to navigate your home using only what it sees.
No LiDAR. No odometry. No GPS. No cloud. Just a camera and a neural network running on your laptop.
Fully autonomous navigation to a goal — learned VQ-VAE perception, CQL policy, zero hand-coded control.
Disaggregated Intelligence
The robot is a thin sensor/actuator relay — a camera, two motors, and a WiFi radio. All cognition runs on your laptop GPU. This decoupling means the robot never becomes obsolete: when better models exist, you update software, not hardware.
The Robot
Camera · Motors · WiFi
Your Laptop
VQ-VAE Perception · CQL Policy · GPU Inference
The robot is a sensor relay. Your laptop is the brain. Neural network complexity has no ceiling.
From Neural Circuits to Neural Networks
Perception — VQ-VAE
A Vector-Quantized Variational Autoencoder compresses each camera frame into a discrete codebook vector — a compact spatial token that captures both where the robot is and which way it's facing. These learned codes are an engineering analog of place cells and head direction cells working together: biological neurons that fire for specific combinations of position and orientation, compressing rich sensory input into a small set of discrete spatial representations.
Action — Policy Network
A Conservative Q-Learning policy takes the current spatial token and a goal token as input and selects the motor action most likely to move the robot toward its target. This is reward-driven action selection — the same fundamental computation that biological neural circuits use to adapt behavior in response to outcomes, learned offline from previously collected experience.
Together, these two networks close the sensorimotor loop: see, encode, decide, act. No hand-coded rules. No LiDAR. No odometry. Just learned spatial representations and learned goal-directed behavior.
What Ships
- XIAO ESP32-S3 Sense — MCU with OV2640 camera
- Custom PCB — Motor drivers, power regulation, sensors
- 2× TT Gearbox Motors — Differential drive
- Laser-Cut Acrylic Chassis — Three-plate sandwich
- IR Proximity Sensor — Forward obstacle detection
- Ball Caster — Rear passive support
Open-source. Full training pipeline. Unbox to autonomous navigation in one session.