Contains expert trajectories tailored for state-only imitation learning tasks in Gym environments.
State-only imitation learning is all about mimicking the behavior of an expert using only the observations of the environment's states. It's like learning to dance by watching someone else's moves without knowing the exact steps/actions they are taking.
In the trajectories directory, you will find the trajectories. Use np.load to use the trajectories for SOIL.
I have also provided the scripts used for collection in the scripts directory. You can change the parameters to collect more trajectories or change the reward threshold or the algorithm used.
| Environment | Algorithm | Details | Trajectories Collected |
|---|---|---|---|
CartPole-v1 |
PPO | I have collected trajectories which yielded a minimum reward of 475. This can be found in the official documentation of the environment. | 1000 |
Pendulum-v1 |
DDPG | Pendulum is an unsolved environment. Hence the approach I have taken is first collect 1e6 trajectories of the trained agent. I have then taken the values that lie greater than mean + standard_deviation of the rewards. I have sampled those trajectories only. |
22 |
MountainCarContinuous-v0 |
DDPG | I have collected trajectories which yielded a minimum reward of -110. | 1000 |