Before anything moves

A policy file drives real motors on a 35 kilogram machine. The sequence below is the one experienced owners follow, and every step exists because someone learned it the expensive way.

1. Match the file to your robot

Check three things against the listing before you buy: your exact platform (a G1 policy does not run on an H1), your hardware revision, and your firmware generation. A mismatch on any of the three is the most common reason a purchased policy disappoints, and the reason Robaku listings are required to state all three.

2. Read the package

A serious policy package contains the checkpoint itself (usually an ONNX or TorchScript file), a configuration file, install and run notes from the seller, and ideally a simulation scene. An action sequence package is simpler: a script written against the platform SDK with its run instructions. If a package arrives as a bare file with no notes, treat that as a quality signal and say so in your review.

3. Simulate first

If the package includes a simulation scene for MuJoCo or Isaac Sim, run it there before the real robot moves. You are watching for three things: does the behavior match the listing video, does it stay within sane joint velocities, and does it recover when nudged. Five minutes in simulation is the cheapest insurance in robotics.

4. First run on hardware, the conservative way

Clear the area of furniture, pets and people. Start from the seller's recommended reduced torque and velocity limits, not from full performance. Keep the remote stop in your hand, not in your pocket. Run the behavior once at reduced limits, watch the whole thing, then step the limits up gradually across runs. Dance and martial-arts sequences deserve extra respect: they use the full workspace of the machine by design.

5. Know what you bought

Licenses in this category are typically per-robot. That means the file runs on your machine, not on your friend's, and reselling the file is not included unless the license says so. Sellers set these terms on their listings; respecting them is what keeps trained-policy creators selling their best work instead of hoarding it.

Where to get them

Browse the category at policies and action sequences, see everything for your platform at the Unitree G1 hub, or if you have trained something worth money, list it.