A humanoid robot, designed in CAD and simulated as designed
The simulation uses the CAD parts with their masses, and the controller sees only what real sensors would deliver.
The robot is built in simulation first, but in a way that gives the result a chance on real hardware. Every joint is modelled as a bus servo, with torque curve, deadband, gear backlash and bus latency.
The controller only sees what real sensors deliver: encoders with their zero errors, a drifting IMU, load cells in the feet, touch pads in the fingertips and the head camera. A test fails as soon as a controller module touches simulator truth. Gait, balance and arms are driven by a learned whole-body controller (PPO). On 2 October 2026 we discarded every result trained with data the real robot does not have, and have been measuring anew since.
- Maturitysimulation, no hardware built
- Size · massabout 780 mm · 4.4 kg
- Drive38 joints: 26 bus servos, 12 finger motors
- Handsfive fingers, six motors per hand
- Controllerlearned (PPO), the robot’s sensors only
- Printed parts107, parametric in CAD
- SimulationMuJoCo, training on the GPU
- Designv7 of 30 September 2026
All figures come from simulation. In tasks that combine seeing, walking and grasping the robot does not yet reach the pass marks, and how much holds on real hardware is open.
Measurements with modelled sensors
Documented state of 5 October 2026. The video shows a newer gait from 9 October for which no measurement series exists yet.
| Check | Scope | Result |
|---|---|---|
| Gait, falls | 108 walks with the head moving | 0 |
| Gait, falls at start-up | 24 starts | 0 |
| Gait, speed | walking and turning | 0.20 m/s · 0.66 rad/s |
| Hand, holding | 512 episodes, thumb and one finger pressing | 373 (73 %), mark 60 % |
The hand learns on its own bench with cylinders, boxes and spheres of 10 to 300 g.
More projects