reproduce the results

[ChenyangRan]

Hi, since the panda-gym cannot set the random seed as gym, where you can use env.set(seed) to reproduce the results.
When I use env.reset(seed=10), if the random seed is the same, I get the same return value, such as desired_goal. But if I don't set the seed, I can't guarantee the consistency of the state of the subsequent env.resets. Is there a way to guarantee the consistency of the state of each subsequent reset just like gym?

[qgallouedec]

Hey, thanks for the question. It seems like a bug, I'll fix it.

[qgallouedec]

I've only managed to solved the issue with some environments. I can't solve it for the others.
The discussion is also ongoing here: Farama-Foundation/Gymnasium#1111

• PandaReach-v3
• PandaReachJoints-v3
• PandaReachDense-v3
• PandaReachJointsDense-v3
• PandaSlide-v3
• PandaSlideJoints-v3
• PandaSlideDense-v3
• PandaSlideJointsDense-v3
• PandaPush-v3
• PandaPushJoints-v3
• PandaPushDense-v3
• PandaPushJointsDense-v3
• PandaPickAndPlace-v3
• PandaPickAndPlaceJoints-v3
• PandaPickAndPlaceDense-v3
• PandaPickAndPlaceJointsDense-v3
• PandaStack-v3
• PandaStackJoints-v3
• PandaStackDense-v3
• PandaStackJointsDense-v3
• PandaFlip-v3
• PandaFlipJoints-v3
• PandaFlipDense-v3
• PandaFlipJointsDense-v3

To reproduce

import panda_gym
import gymnasium as gym
from gymnasium.utils.env_checker import data_equivalence


env = gym.make("PandaPickAndPlace-v3")

action_0 = env.action_space.sample()
action_1 = env.action_space.sample()

_, _ = env.reset(seed=10)
obs_00, _, _, _, _ = env.step(action_0)
_, _ = env.reset()
obs_01, _, _, _, _ = env.step(action_1)

_, _ = env.reset(seed=10)
obs_10, _, _, _, _ = env.step(action_0)
_, _ = env.reset()
obs_11, _, _, _, _ = env.step(action_1)

assert data_equivalence(obs_00, obs_10) 
assert data_equivalence(obs_01, obs_11)