Rocky is an open-source, four-legged robot with all of its sensing, compute, and power onboard — nothing offloaded to a phone or a server. A deliberately modular base for experimenting with locomotion and agentic robot behavior in the home.
Sensing, compute, and power all ride on the robot. No phone, no server, no tether — it works on its own.
Twelve identical actuators run a learned policy — an efficient dog-like trot, or a sprawled, omnidirectional crab walk.
A vision-language model reads the scene and your request, then dispatches structured navigate, look, and walk commands.
Maps the home once, then localizes against that map with stereo depth, an IMU, and SLAM running continuously.
A microcontroller keeps it on its feet in real time while the Jetson handles the heavier thinking over ROS 2.
Swap actuators, sensors, and compute. Weight- and torque-conscious throughout — every part is a real tradeoff.
Drag to look around, and switch between the two gaits. Same hardware, two learned ways to move — nothing else is attached, and nothing it depends on lives off the robot.
Robotic arms have solid open platforms to build on — projects like OpenArm and YAM give researchers a shared base. Legged locomotion doesn't really have that yet, and neither does the harder problem stacked on top of it: moving through and understanding an indoor space well enough to live alongside people.
Navigation and spatial intelligence to co-exist with humans in the home is just as big a problem as manipulation — and far less explored.
Rocky is an attempt at that missing base. Actuators, sensors, compute, and frame are chosen to match each other — and meant to be swapped. This is the foundation; the interesting part is what gets built on it.
Rocky is always listening and always knows where it is. Say something, and the request goes to a vision-language model that looks at what the camera sees, decides what to do, and turns it into a clear instruction. The navigation stack and the gait policy carry it out, while a voice and a small animated face respond.
Two loops run in parallel: the microcontroller handles walking — the fast loop that keeps it upright — while the Jetson does the slower, heavier thinking over ROS 2. The dog-like and crab-like gaits are just different policies on that same real-time loop.
Everything has to ride on legs that can still move it, so each part is a tradeoff between capability and weight. The full bill of materials, grouped by subsystem.
| Motion | |
| DAMIAO DM-J4310-2EC12 × leg actuator | Clean CAN smart servos, low backlash, good torque-to-speed balance. |
| Perception | |
| Orbbec Gemini 336 | Stereo depth + IR + IMU, sized for house-scale depth. |
| Dynamixel XL430-W250-T2 × pan / tilt head | Position feedback, reliable, easy to mount. |
| Ultrasonic rangefinders | Close-range safety net for blind spots and glass. |
| Intelligence | |
| Jetson Orin NX 16GBSeeed reComputer J401 | The ROS 2 brain — USB3, CAN, GMSL2, UART, I²C. |
| STM32H7+ 2 × isolated CAN | Real-time motor-bus control and safety. |
| Power | |
| 6S2P high-current Li-ionP50B / P45B cells | Better weight-to-runtime than LiPo. |
| Fused PDB + 12V & 5V bucks | One clean battery system with regulated rails. |
| Voice & face | |
| ReSpeaker Mic Array v2.0 | Far-field voice pickup and rough sound direction. |
| Speaker + MAX98357A amp | Text-to-speech out — gives it a voice. |
| Small OLED / LCD | A face for status and expression. |
Component choices are still settling — this is a working base, not a frozen spec.