Question regarding the capability of the simulator about rlbench HOT 2 CLOSED

Hi,

• Yes, this is a typo in the documentation. It should read: "As previously mentioned, the physics engine runs far faster and smoother when all shapes for which it makes calculations are non-convex.". I'll update this in version 1.0.6.

• I did not actually write the tutorial, so not exactly sire what is trying to be conveyed here. But regardless, it doesn't seem important. I will go through the tutorials for version 1.0.6 and try and clean them up.

• The 'fake' grasping behaviour is done by using a proximity sensor that is located in the gripper. If at any point a registered 'graspable' object is located proximity sensor when a close gripper command is sent, then the gripper will close until hitting some maximum force, and the object will be attached as a child to the gripper. To answer your specific questions:

  • yes a larger objects could in theory be grasped. This is a trade off between realistic grasps and task feasibility.
  • No it is any point on the object, not the centroid. (this is how the proximity sensors work).
  • The angle does not matter

• Yes there objects that are assumed to be static (i.e. respondable but not dynamic). Yes there are joint limits; so in the door case, the door joint can only range from 0 to ~ 40 degrees. The rope is modelled via a series of spherical joints connected via dynamic objects that are able to intersect with each other (using collision masks). Yes there are sprint constraints in CoppeliaSim; you can set a joint to mass/spring mode via the joint dynamics properties, where you can directly specify the spring constant and the damping coefficient. I will take your door 'tutorial' feature request on board, but am quite busy with general maintenance and my own research, so cannot give a timeline sorry.

Will close, but feel free to re-open if you have any other questions.

Best,
Stephen

from rlbench.

Is there a possibility of overriding the fake grasping behavior to give more realistic results?

from rlbench.