Seems like we have a bug somewhere in the pybullet simulation code. The pointcloud is not aligned with the world. My guess is that the transform that PyBullet is using to render the camera image is somehow not matching up with the one being published out to the ROS tf server.