Why different dimensionality state tensor with n_obs_steps vs not?

[genemerewether]

Curious about a design decision - why not have ACT with a [batch, n_obs_steps, state_dim] tensor but assert that n_obs_steps is length 1? Instead of [batch, state_dim]

Currently, we have to detect different dimensionality and handle when we're writing policy-agnostic code

[genemerewether]

E.g. #194 - just adding the delta_timestamps changed the state dimension