All policy should use relative actions

[xianglunkai]

I seem lerobot and openpi use relative actions for training and inference, and it would improve performance of all policy.

[shuheng-liu]

Can you clarify what "relative action" entails here?

• If you meantaction[t] - action[t0] for t in {t0, t0+ΔT, ...}, then the policy still needs to produce a action[t0] in the forward pass.
• If you meant action[t] - state[t0] for t in {t0, t0+ΔT, ...}, then you are effectively assuming states and actions share the same dimensionality, which is not the case for every LeRobot dataset.

[xianglunkai]

@shuheng-liu
Txs!
I meant action[t] - state[t0] for t in {t0, t0+ΔT, ...} . From real robot experiments, relative actions could have more high success rate and smothering performance.

[shuheng-liu]

Implementing it should be easy but the problem is we can't assume all lerobot dataset has action and state in the same space. For example, observation.state is 8-dimensional and action is 7-dimensional in the lerobot/libero dataset