A policy that teleports the agent to a goal, and takes it back if that lands somewhere useless.

Mirrors tbp/monty/frameworks/models/motor_policies.py::JumpToGoal, the policy a distant agent's selector reaches for when a learning module has a hypothesis worth testing. A jump is two actions — SetAgentPose at the goal's location facing along the goal's direction, and an identity SetSensorRotation so the sensors keep the orientation the jump was planned in — and it always remembers the two actions that would put the agent back where it started.

It is meant to be called twice: once to jump, and once to judge where the jump landed. Landing on the object is the end of it, and the selector moves on to the next policy; landing anywhere else — empty space, or inside something — answers with the undo actions instead. The judgement reads the percept (on_object), not the depth, which is where it differs from InformedPolicy's version of the same idea.

With no jump in flight and no goal it raises NoGoalProvided, because a selector that routes a goal to this policy has already decided there is one; a run configured to suppress runtime errors gets a warning and no action instead.

:sensor_id is stored and never read, as in Monty, whose docstring says it is for a depth-at-centre calculation this policy no longer does.