What's this about?
Pollen open-sourced the full training pipeline. The robot learns in MuJoCo using PPO with GPU-based training; the trained policy is exported to ONNX and deployed to the physical robot. The part that makes it work is what the simulator bothers to model: actuator behaviour, battery voltage, command delays, friction — the unglamorous physics that normally makes a simulator-trained policy fall over on real hardware.
What is "sim-to-real"?
Training a robot in the real world is slow and destructive — you get one attempt per second and you break the hardware. Training in simulation gives you millions of attempts in parallel with nothing to repair. The catch is that the policy learns the simulator's physics, not the world's, so it fails the moment reality differs. Sim-to-real is the set of tricks for closing that gap, mostly by deliberately making the simulator imperfect and varied so the policy has to be robust rather than precisely tuned.
Why is it worth noticing?
- The full stack being Apache-2.0 means the pipeline is reusable for hardware that isn't a duck.
- Modelling battery voltage is the detail that signals this is a real deployment pipeline, not a demo — policies degrade as the battery drains, and most simulators ignore it.
- MuJoCo + PPO + ONNX is the boring standard stack, which is a good sign for anyone wanting to reproduce it.