Yes, this type of controller has two purposes once Control Barrier Functions are included: (1) It can steer agents to desired goals. (2) It can be used as a safety shielding (For example, to avoid collisions or never exceed certain relative velocities with respect to other agents). With (2), the controller can take an action from another policy and use it as a reference. The safety shielding will only be activated if the safety constraints are violated. This feature can be used for safe RL learning.

Yes, this type of controller has two purposes once Control Barrier Functions are included: (1) It can steer agents to desired goals. (2) It can be used as a safety shielding (For example, to avoid collisions or never exceed certain relative velocities with respect to other agents). With (2), the controller can take an action from another policy and use it as a reference. The safety shielding will only be activated if the safety constraints are violated. This feature can be used for safe RL learning.

Ok so here is what we can do. As I understand it, this heuristic provides an example of (1). We can merge this as a heuristic for navigation as it is specific to goal following.

Regarding (2), it would be nice to have a customizable component in https://github.com/proroklab/VectorizedMultiAgentSimulator/tree/main/vmas/simulator/controllers that takes some custom constraints and applies the CBF to the agent action. This compenents will then be called in scenario.process_action()

A challenge here is how to create such a component so that it is able to take general constraints.