RoboticsPart 2 of 2Drawing 3.2 of 37
Learning to walk
A policy I've trained in simulation since August 2026. This part covers what it can do, the three things that broke getting there, and why none of it has been on the robot.
- StatusComplete, September 2026. Robot built 2021 to 2022, software rebuilt from scratch from August 2026
- Joints12 servos, three per leg, 1.96 N.m each
- Mass3.1 kg
- BrainRaspberry Pi, PWM at 50 Hz, three IMUs on the trunk
Sheet 02 of 04/Walking
Teaching it to walk
The robot's apart on a shelf right now, so everything below comes from simulation, with 64 robots running at once.
| Criterion | Bar | Result | Policy |
|---|---|---|---|
| Walks forward | 5.0 m | 6.7 m | Flat ground |
| Top speed | none set | 0.72 m/s | Flat ground |
| Holds a straight line | 7 deg | 4.5 deg | Flat ground |
| Stays upright | 100 of 100 | Flat ground | |
| Climbs a 10 deg slope | 98 of 100 upright | Slope | |
| Walks sideways | short by 3 mm/s | Flat ground | |
| Criteria passed | 11 | 10 | Neither alone |
One policy trained on flat ground made most of this table, and a second one trained on slopes made the climb. No single policy has done both yet.
Sheet 03 of 04/What broke
What I got wrong
I built the software stack on a bad CAD export, so every policy trained on it was learning to walk as a robot that didn't exist. In August 2026 I archived all of it to a git tag and started over from SolidWorks.
Sheet 04 of 04/Next
Where it goes next
Right now the policy trots, so the body falls between steps and the servos catch it twice a stride. I don't think hobby servos can take that on the real robot, so this version stops here. The learning setup moves on to Gray 2.0.