Gump Robotics builds small, agile two-legged robots that walk around your table, learn tricks and keep an eye on your coffee. F1 is the first.
Joint test, from our design filesRight hip pitch
You don't program F1, you teach it, the way you'd teach a dog. These are the things we're teaching it first.
Walk and trot across the table, quickly and without falling off the edge.
Every now and then, walk over to your cup, look inside and tell you when it's empty.
Look around and tell you what's lying on your desk, and where you left your keys.
Gymnastics moves and party tricks that you teach it yourself, with rewards instead of code.
One day, on a tiny skateboard. It's our hardest goal, and the one that keeps us honest about agility.
Capable humanoids exist, but they are priced for labs and factories. The affordable small ones mostly shuffle. Gump sits in between: the engineering of a research robot, at the size and price of a toy you'd actually buy.
Agile and capable, from around US$16,000 (Unitree G1) and up. Built for labs.
Affordable, but slow and stiff. They shuffle rather than walk and can't really balance.
Built for dynamic balance, learned walking and real sensing, in a 28 cm body you can afford.
Every part of F1 is designed in-house, tested in simulation first and then built. Here's what goes into it.
Each foot tilts forwards and sideways, driven by two servos through tie rods with ball joints on the outside of the leg. Both servos share the load, so the ankle is twice as strong.
A physics model of F1 in MuJoCo, built from the CAD and the measured servo strength. New behaviours are tried there thousands of times before the real robot moves.
A neural-network policy trained with reinforcement learning in the twin runs on F1's own microcontroller, correcting every joint 50 times a second.
A 3D-printed 3:1 planetary gearbox, designed and tested for the hips, triples the torque where balance needs it most.
| Height and mass | About 28 cm and 0.5 kg |
|---|---|
| Joints | 12: hip pitch, roll and yaw, knee, ankle pitch and roll on each leg |
| Actuators | 12 digital servos with position feedback |
| Brain | ESP32-S3 controller, plus a second processor that reads every joint's position |
| Sensing | 6-axis motion sensor (gyro and accelerometer), joint positions |
| Control | Whole-leg kinematics, ankle balance from the motion sensor, load compensation, learned walking policy |
| Body | 3D-printed structure with a vented electronics enclosure |
F1's ankle rods sit on the outside of each leg, running from the calf down to the foot. When the first prototype stood up, it looked a lot like young Forrest in the film Forrest Gump, walking in his leg braces.
The name stuck. The company became Gump Robotics, and the first robot became F1, for Forrest 1. Like Forrest, we expect it to start a little wobbly and then run.
Built step by step, simulation first. Where F1 is today:
Designs, builds and programs F1: the mechanics, electronics, firmware, simulation and learning.