I am currently reading the paper Learning agile and dynamic motor skills for legged robots. The paper claims that robotic behavior learned in simulation is difficult to transfer to a real-world situation as a realistic simulation is difficult. The authors, therefore, enrich a typical simulation substantially. To quote:

Our key insight on the simulation side is that efficiency and realism can be achieved by combining classical models representing well-known articulated system and contact dynamics with learning methods that can handle complex actuation. The rigid links of ANYmal, connected through high-quality ball bearings, closely resemble an idealized multibody system that can be modeled with well-known physical principles (40). However, this analytical model does not include the set of mechanisms that map the actuator commands to the generalized forces acting on the rigid-body system: the actuator dynamics, the delays in control signals introduced by multiple hardware and software layers, the low-level controller dynamics, and compliance/damping at the joints. Because these mechanisms are nearly impossible to model accurately, we learned the corresponding mapping in an end-to-end manner

I am asking myself how to represent these things in a Gazebo simulation. In particular:

1. Are the rigid links represented by the collision, visual, and inertial properties in a URDF file?
2. How are the torque commands represented in Gazebo? Do they correspond to the tags inside a joint's description in a robot's URDF file?
3. Is it possible to refer to an external program in Gazebo to calculate the torque forces during the simulation based on some publications from ROS instead of specifying these things inside a URDF file?