VelStand — walking and fall recovery together

Mjlab-VelStand-Flat-MicroDuck
Mjlab-VelStand-Rough-MicroDuck

Walking and fall recovery trained into one policy.

Why not just train them separately

The runtime can already hot-swap between policies — all of them share the 61-dimensional observation contract, so any one can take over at any moment. Walking and standing up can perfectly well be two policies.

So why train them together?

A combined policy is more coherent through the transition from recovered to walking, because it has seen that transition. Two separate policies are each excellent, and neither has specifically practised the handover.

The cost is that it is harder to converge — one policy learning two quite different things.

Which to choose

If you want the robot pushed over, back up, and walking again without a seam, take this. If you want each behaviour trained to its best and can accept a slightly abrupt switch, use Velocity plus StandUp.