[Question] Stochastic of environmental dynamics in Gym control tasks

[return-sleep]

Question

When I download the offline dataset and want to reproduce the trajectory with the same action sequences and initiate states, the subsequent state sequences (obs &reward) gradually offset the original offline trajectory over time.

test1.mp4

The following code is used.

dataset = env.get_dataset()
states = dataset['state'][0:H] # we sample the first trajectory
actions = dataset['state'][0:H]
init_state = states[0]
env.set_state[init_state]

for t in range(H):
      obs, reward, done,info = env.step(actions[t])

I hope for your reply!

[EdanToledo]

I've also noticed this! Any update?