clvrai/mopa-pd

Distilling Motion Planner Augmented Policies into Visual Control Policies for Robot Manipulation (CoRL 2021)

mopa-pd's Issues

Keyword is_bc_policy

Hi!

I was trying to go through the training process with SawyerPushObstacle-v0. I met the keyword error when the following line is executed

mopa-pd/rl/behavioral_cloning_visual.py

Line 367 in ac55f56

obs, reward, done, info = self._env_eval.step(action.detach().cpu().numpy(), is_bc_policy=True)

where the step function in BaseEnv has no is_bc_policy keyword.

mopa-pd/env/base.py

Line 264 in ac55f56

def step(self, action, is_planner=False, is_mopa_rl=True):

Or, the _step method should be called here?

mopa-pd/env/sawyer/sawyer_push_obstacle.py

Line 192 in ac55f56

def _step(self, action, is_planner=False, is_mopa_rl=True, is_bc_policy=False):

Looking forward to your reply, thanks.

About performance in the paper

Hi,
I have some problems about the performance in the paper.
In [Table 1], which methods were trained with MoPA-RL trajectories?
I have seen that MoPA-RL trajectories were used by BC-Visual and Ours.
Can u explain the training data for each methods, i.e., Asym. SAC | MoPA-Asym. SAC | CoL | CoL w/ BC smoothing | Ours w/o BC smoothing?
I am confused that why [CoL w/ BC smoothing] cannot perform well if it was also trained with MoPA-RL trajectories and BC smoothing, just as your method MoRL-PD?

Thanks for your any reply.