Closed Jialn closed 5 years ago
Also added rendering mode "rgb_array". Since gzclient might not be always opening, and we can not directly access user camera of gzclient in gzserver, this mode is done by inserting a camera to the world.
The default view pose is like below. User can set view point by setting pose of "self._rendering_camera" in "env.step()" .
Does alf.trainer.on_policy_trainer.play() generate right video from this?
It seems that you need to add metadata like this so the video can be captured: https://github.com/bulletphysics/bullet3/blob/0aaae872451a69d0c93b0c8ed818667de4ad5653/examples/pybullet/gym/pybullet_envs/env_bases.py#L16
It seems that you need to add metadata like this so the video can be captured: https://github.com/bulletphysics/bullet3/blob/0aaae872451a69d0c93b0c8ed818667de4ad5653/examples/pybullet/gym/pybullet_envs/env_bases.py#L16
OK, I'll add it.
Because the rendered image has no grid, and the yellow texture in texture_ground_plane draws too much attention, I changed the yellow texture to grey texture with 1m x 1m grid:
the default view point, this also suits for icubwalk: PR2 has been set to a closer camera pos simple navi:
Do the model trained with the previous ground plane texture work with the new ground texture?
And since the ground texture is changed. Perhaps we should update the videos in README.md
Some minor changes like using action_space.sample() instead of np.random(action.shape), use np.clip instead of min(max()), etc.