Kuka robot arm pybullet env doesn't render?

[maximecb]

Hi. I am not seeing any visual output for the kuka env, when calling env.render('human').

Test code to reproduce:

import gym
import pybullet_envs
import numpy as np
import time
env = gym.make('KukaCamBulletEnv-v0')
env.reset()
env.render('human')

for i in range(50):
    action = np.random.uniform(low=-1, high=1, size=(env.action_space.shape[0],))
    obs, reward, done, info = env.step(action)
    time.sleep(0.2)
    env.render('human')

The same code works for the reacher environment, with only the environment name changed.

[maximecb]

Upon further examination, the reacher env does produce visual output, but the camera is very far. There is a method camera_adjust defined here, but I think this method is never called, because when I call it, there seems to be a bug in the method:

  File "<stdin>", line 1, in <module>
  File "/home/maximecb/.local/lib/python3.6/site-packages/pybullet_envs/gym_manipulator_envs.py", line 44, in camera_adjust
    self.camera.move_and_look_at(0.3, 0.3, 0.3, x, y, z)
  File "/home/maximecb/.local/lib/python3.6/site-packages/pybullet_envs/env_bases.py", line 138, in move_and_look_at
    self._p.resetDebugVisualizerCamera(distance, yaw, -20, lookat)
AttributeError: 'Camera' object has no attribute '_p'

[maximecb]

I am able to correct the camera position in the reacher environment manually like this:

import pybullet
pybullet.resetDebugVisualizerCamera(0.5, 0, -45, (0, 0, 0))

[erwincoumans]

Can you call env. render('human') once before the first env. reset? The issue is that PyBullet needs to create an OpenGL window at initialization. Will be fixed in a future major revision.

[maximecb]

@erwincoumans I still get no window showing on the kuka env, whether I call env.render() before or after reset.

[maximecb]

Ping @erwincoumans. Can you check that you can get the kuka env to show a render window?

[Teslatic]

Hey @maximecb,
I stumbled on this issue as well and found a workaround. Maybe it comes a little bit late, but maybe it can help others:

from pybullet_envs.bullet.kukaGymEnv import KukaGymEnv
from itertools import count

env = KukaGymEnv(renders=True, isDiscrete=False)
env.reset()

for t in count():
    action = env.action_space.sample()
    next_obs, reward, done, _ = env.step(action)

    if done == True:
        break

    obs = next_obs

[swagatk]

Dear @maximecb, thanks for your code. It does work and I am able to see the simulation environment. Does this mean that I can now implement my RL algorithm on this? Where can I find more information about various action and observation fields used by this simulation? Is there any documentation available for this environment? I am using the Cam version of this environment: KukaCamGymEnv.

[zichunxx]

Hi! If all the Kuka-related reinforcement env won't be updated for the new gymnasium API?