InnerModel and PyBullet

In this post we will talk about the features needed by the InnerModel lib and features provided by the PyBullet lib. We will analysis the kind of control that the PyBullet lib provides, how many different ways we can render something, how a multilink body is organised in its APIs, how the compromises have to be made because of some features which pybullet can not provide

• As I had already described in the last post, InnerModel lib is C++ library, used to represent an innermodel scene describes as an xml file. InnerModel C++ uses openscene graph lib for rendering all the graphics

• The features that innermodel C++ provides:
  • control of the robot through joints
  • giving access to the camera from different points of view
  • providing access to a debug window
  • point cloud
  • lasers
  • moving a body
  • joint control
• So let me talk one by one how the APIs provide PyBullet lib achieves these requirements

  • One of the first challenges we face is how to represent a body, because in PyBullet only after creating a body one can render anything in PyBullet. The other way is to feed a urdf file, which is completely different case. createMultiBody() is the API used to form a body. It needs various attributes to render the body but two of the most important ones are collisionShapeIndex and visualShapeIndex which is created by createVisualShape() and createCollisionShape() which takes in 3D mesh files are argument.

  • So from above we can infer that creating a body is the first and the most important thing we require in PyBullet. So how do we read separate different bodies from an xml file. The manner in which xml for innermodel is written is particularly different from the likes of ROS’s urdf or sdf where there are tags which differentiate between parent link and child link. So some thought has to be given before reading and writing the xml. The way it is done now is by considering the children of innermodel parent tag (<innermodel>) or transform tag with “world” as id (<transform id="world">), at the beginning of the file. Beginning of a body is always a transform tag (<transform id="name">) with name of the body as its id. After we have declared the container for the whole body, what can we do with it. So inside the body tag we can have mesh tag (<mesh>) to include mesh files or for simple cuboidal shape we include planes (<plane>). To specify transformation between the different component of a robot we use transform tag (<transform>). For including any sensor in the robot we just put a tag of that name (<laser>, <camera>, etc).

  • One of the most important sensor that a simulator has to provide is camera. PyBullet provides getCameraImage() and getDebugVisualizerCamera() for getting images from the simulator. getDebugVisualizerCamera() is used to get default GUI that all the simulators give. While on the other hand getCameraImage() is used to get from any point of view. The API require camera speci- fications like focal length, near and far distance, width and length of the image. while on the other hand getDebugVisualizerCamera() doesn’t require any requirements at all.

  • Next is point cloud. PyBullet doesn’t have anything direct for point cloud but we can have a work around which creates a very coarse point cloud but point cloud all the same. We use the position of the camera and where the light rays from the camera might intersect with the objects in the simulation to get a depth image and from that depth image we can render tiny spheres in the GUI to make it appear like a point cloud

  • For lasers, PyBullet does have an API which lets emulate laser rays. rayTestBatch() is an API which takes two arrays: from & to denoting source and desitnation of laser rays and outputs the 3D points intersection points. These rays can be further visualized using addUserDebugLine() which can render a line of any color from a source and a destination.

  • The innermodel lib can also support IMU and that can be easily implemented in PyBullet as it has an API getBaseVelocity() which outputs velocity of a body, from which we can derive the accln.

  • The next most important thing we want in a simulator is control. How can we change position of a body? There are two ways we can exert control on a body in PyBullet, either move the whole body or just control the joints of the body. resetBasePositionAndOrientation() resets the position and orientation of the base of a body taking two arguments of reset position and reset orientation. The other way through joint is by setJointMotorControl2() through which we can set target position target velocity and force as arguments which can also vary the parameters of position gain, velocity gain, etc.