Adroit
A two-legged robot that drives on wheels and worked out how to balance on its own.
- Year
- 2026
- Type
- Wheeled biped
- Balance
- Reinforcement learning, Isaac Lab
- Onboard
- Jetson Orin Nano
FIG. I ASSEMBLY · FUSION 360 
FIG. II THE BUILT ROBOT · 2026 What it is
Adroit is a wheeled biped: two legs, with a powered wheel at the end of each. Wheels are fast and efficient on flat ground. Legs let the robot change its height and get over things a wheel on its own would catch on. Putting both on the same machine means it has to balance, and balancing is the interesting part.
It was drawn in Fusion 360 first. The render above is the assembly: a capsule body holding the electronics, a linkage in each leg, and treaded wheels at the ends. The photo beside it is the same design once it was built.
Learning to balance
Balancing on two wheels can be written out as control equations, or the robot can work it out for itself. Adroit does the second. Its balance policy is a neural network trained by reinforcement learning in NVIDIA Isaac Lab, where copies of the robot run in parallel and keep falling over until they stop. Falls in a simulator cost nothing, which means the useful ones can happen there instead of on real hardware. I named the robot Adroit after the princeton compute cluster Adroit, that allowed me to run the training on powerful A100 GPUs
Running onboard
The trained policy is deployed to a Jetson Orin Nano that rides on the robot, so it balances by itself rather than through a laptop or a tether. This is the part where simulation gets tested honestly. The motors have backlash, the floor is never quite level, and the sensors are noisier than anything in Isaac Lab. Nothing ever goes to plan...