Policy hot-swapping

Microduck’s runtime hot-swaps between policies: walking, fall recovery and each skill is its own .onnx, loaded by robotd on demand.

It rests on the shared contract

This works because every policy shares one 61-dimensional observation contract. Switching is loading a different network — no state translation, no reshaping.

A concrete case: walk ⇄ roller

After fitting or removing the wheels, hold D-pad up for three seconds:

1. The robot quacks once for walking, twice for roller
2. Returns to its home pose
3. Loads that mode’s policies there
4. Drives again

A few seconds, torque on throughout, no restart.

robotd.toml is not touched, so a reboot comes back in the configured mode. Make it stick with robotctl configure and [policy] mode.

It is a hold rather than a press because D-pad up is easy to lean on while driving.

Running your own

No release needed. Point the config at your file:

[policy]
walk = "/home/radxa/my_walking.onnx"
stand = "/home/radxa/my_stand.onnx"

sudo systemctl restart robotd

Those paths survive updates — a release replaces what it ships, not the file pointing elsewhere.

A policy that could not be loaded reports unhealthy, and both robotctl health and the bottom border of monitor name the reason.