vitchyr / multiworld Goto Github PK

in this file

line 51 to 54

        if puck_low is None:
            puck_low = self.hand_low[:2]
        if puck_high is None:
            puck_high = self.hand_high[:2]

should we init puck_low/high with paras in line 15 to 16 ?

            puck_low=(-.4, .2),
            puck_high=(.4, 1),

as title

A config option to deactivate the teleportation notification would be nice. Thank you! :)

Hi there,

I want to ask that, is there some tricks to set some keys like proprio_observation, proprio_desired_goal, proprio_achieved_goal. Also, are those things redundant for training?

Cause, when i only keep:
state_observation=flat_obs,
state_desired_goal=self._state_goal,

some experiment based on this environment could not converge with soft a-c, could you help explain a little bit about this?

Thanks

Hi,

Maybe my question is simple! I am trying to test RIG method. At first I made a conda environment and then I clone Multiworld repository to the Conda packages. When I run a the python program, it shows that SawyerPushNIPS-v0 is not registered. I check the codes, and I think there is not any connection between gym and Multiworld.
How can I solve this issue?

Thank you very much

In generate_uncorrected_env_goals, the return block needs to be unindented by one level or it doesn't return anything when fix_goals == True.

Hi there,
I am trying to use Sawyerxyz environment with sawyerpushandreach task. I found that when i load a mesh as puck, the robot could push the object without "real contact" with the puck(but when enter c from keyboard ,it shows contact on the air not on the object).
Also, the gripper sometimes has abnormal actions
Could you help solve this problem? Thanks

The cameras in camera.py seem to only change the viewer viewpoint. What if I want to change the camera viewpoint of the images outputted in line 177 of image_env.py?

        image_obs = self._wrapped_env.get_image(
            width=self.imsize,
            height=self.imsize,
        )

wrapped_env.get_image calls self.render(..., camera=None). It doesn't seem to use the camera loaded into the viewer from cameras.py.

In the sawyer reaching and pushing environments, the reference to the sawyer_control seems to be missing.

After

git clone [email protected]:vitchyr/multiworld.git
cd multiworld
pip install -e .
python  # version 3.8.0
>>> import multiworld

I get

ModuleNotFoundError: No module named 'gym'

After i pip install gym, I get

ModuleNotFoundError: No module named 'pygame'

Hello,

I've been working with this repo for the past week, and have noticed (and as mentioned in the documentation) that the sampled goals where the puck is in the hand in sawyer_pick_and_place.py, an uncorrected hand-object configuration is randomly sampled, evaluated, and the final corrected hand-object config is stored in self.presampled_goals.

As you all mentioned in set_to_goal, the hand-object config stored in self.presampled_goals can often fail (the object often does not show to be in the hand at the end of the grasp). I made a simple heuristic for checking whether an attempted grasp was successful (seeing if the object is at least slightly above the table), but I wanted to store the uncorrected hand-object configuration for successful grasps, rather than the corrected one.

My question is, will storing the "wrong", uncorrected hand-object configuration returned from generate_uncorrected_env_goals() into self.presampled_goals interfere with any other parts of the project? While the compute_rewards method in sawyer_pick_and_place.py is using the sampled goals directly, the SkewFit algorithm should just be using latent distance rewards. Does the VAE by any chance use the goal positions in any way? I figure the compute_rewards method in sawyer_pick_and_place.py is being used only for logging anyway.

I'm just trying to run SkewFit by the way. I'm not interested in any of the other RL algorithms that could be using the goal positions. Thank you ahead of time!

I'm not sure if I'm installing this wrong but I clone the repo, cd into the repo and do pip install -e .. However, when I do this, many subdirectories are not available for import? More details here:

https://stackoverflow.com/questions/64143615/pip-install-is-not-installing-all-subdirectories

Hi! I was trying to use the keyboard control script to control the agent, but pressing the keys for movement (mentioned in the file) seem to have no effect. What am I doing wrong?

Thanks!

I ran the code from the rlkit ,but it seems that something is missing in the environment.

Traceback (most recent call last):
File "rig/pusher/rig.py", line 86, in
use_gpu=True, # Turn on if you have a GPU
File "/home/ubuntu/rlkit/rlkit/launchers/launcher_util.py", line 166, in run_experiment_here
return experiment_function(variant)
File "/home/ubuntu/rlkit/rlkit/launchers/rig_experiments.py", line 40, in grill_her_td3_full_experiment
train_vae_and_update_variant(variant)
File "/home/ubuntu/rlkit/rlkit/launchers/rig_experiments.py", line 86, in train_vae_and_update_variant
return_data=True,
File "/home/ubuntu/rlkit/rlkit/launchers/rig_experiments.py", line 248, in train_vae
variant['generate_vae_dataset_kwargs']
File "/home/ubuntu/rlkit/rlkit/launchers/rig_experiments.py", line 166, in generate_vae_dataset
env = gym.make(env_id)
File "/home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/gym/envs/registration.py", line 167, in make
return registry.make(id)
File "/home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/gym/envs/registration.py", line 119, in make
env = spec.make()
File "/home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/gym/envs/registration.py", line 86, in make
env = cls(**self._kwargs)
File "/home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/multiworld-0.0.0-py3.5.egg/multiworld/envs/mujoco/sawyer_xyz/sawyer_push_nips.py", line 478, in init
**kwargs
File "/home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/multiworld-0.0.0-py3.5.egg/multiworld/envs/mujoco/sawyer_xyz/sawyer_push_nips.py", line 59, in init
MujocoEnv.init(self, self.model_name, frame_skip=frame_skip)
File "/home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/multiworld-0.0.0-py3.5.egg/multiworld/envs/mujoco/mujoco_env.py", line 28, in init
raise IOError("File %s does not exist" % fullpath)
OSError: File /home/ubuntu/anaconda3/envs/rlkit/lib/python3.5/site-packages/multiworld-0.0.0-py3.5.egg/multiworld/envs/assets/sawyer_xyz/sawyer_push_and_reach_mocap_goal_hidden.xml does not exist

Hi there,

I am using sawyer_pushing under envs/realworld/sawyer folder and is wrapped by ImageEnv. When I reset, the ImageEnv will perform as follow:

In if-else case, it will jump to 3rd case (L144-L147). I this situation, I understood that you perform as follows:
Step 1: backup the current state (L144),
Step 2: sample and move to the goal state (L145)
Step 3: get the image observation of goal
Step 4: return to the original location (L147)

At step 2 & 3, in sawyer_control, I saw it require the user to move the object into a goal location to capture the image (link below). But at step 4, I didn't see that it requires the user to return object in reset location. In the simulator, everything is ok.

https://github.com/mdalal2020/sawyer_control/blob/3df2c9e246240388bb2bb5f51ecbfd558a1ef74c/src/sawyer_control/envs/sawyer_pushing.py#L33-L39

Is there any missing here?

I am not sure should I post the issue here or sawyer_control? So if it is not suitable, I will move to sawyer_control.

Thank in advance!

I was wondering where I can install the railrl referenced from in the multiworld/multiworld/envs/pygame/walls.py file?

from railrl.torch import pytorch_util as ptu

The mocap for the hand doesn't line up with the actual position of the hand. Thus, if we do something like

env = gym.make(some_sawyer_env_id)
obs_dict = env.reset()
print(obs_dict['state_achieved_goal')
obs_dict, rew, done, info = env.step(0*env.action_space.sample())
print(obs_dict['state_achieved_goal')

the second obs_dict should be the same as the first since we didn't do any action. However, since actions set the mocap position relative to its initial position, and since its original position does not line up with what is displayed in the initial obs_dict, the resulting position after "doing nothing" is significantly different from where it starts.

Is there an easy way to get the unflattened image observation (from Mujoco for example)? And I can't seem to get the same image as the one that shows up in render when I use the internal get_image function:

image_obs = self._wrapped_env.get_image(
            width=self.imsize,
            height=self.imsize,
        )

Is there a way to just get the exact same image that mujoco renders, except as the image observation itself?

Hi, there. I find a compatibility issue.

According to openai/gym@a99e8d1, the latest version of gym has remove the argument tags in the registration function, which makes it incompatible with the registration process in this repo.

I suggest to add gym<0.16 to setup.py to tackle this issue. I have tested it on gym==0.15.7, and everything works fine.

I tried to run "SawyerPushNIPS-v0" but I have the following error:

ERROR: GLEW initalization error: Missing GL version

I installed mujoco-py on my computer and it works well. Anyway to fix it ?