Robot navigation using a high-level policy model and a trained low-level policy model

A method of navigating a mobile robot in an environment, the method comprising:

• identifying a target label for a navigation target in the environment;
• navigating the mobile robot to the navigation target, navigating the mobile robot to the navigation target comprising, at each of every iteration during navigation to the navigation target:
• obtaining corresponding current observation data that is based on corresponding current output from a sensor component of the mobile robot, wherein the corresponding current output from the sensor component is a first type of observation;
• processing the corresponding current observation data, and the target label, using a trained high-level policy model to generate a corresponding high-level action output;
• using the corresponding high-level action output to select a corresponding particular high-level action from a plurality of discrete high-level actions of a defined high-level action space;
• determining whether the corresponding particular high-level action is one that is capable of resulting in collision;
• when it is determined the corresponding particular high-level action is not one that is capable of resulting in collision:
• controlling one or more actuators of the mobile robot based on a corresponding default low-level action defined for the corresponding particular high-level action;
• when it is determined the corresponding particular high-level action is one that is capable of resulting in collision:
• using a trained low-level policy model to generate, based on processing current additional observation data and the high-level action, corresponding low-level action output that is based on the high-level action and that is optimized, according to the low-level policy model,
• wherein using the trained low-level policy model to generate the corresponding low-level action output is without utilizing any observations that are of the first type of observation, and
• wherein the current additional observation data is based on corresponding current additional output from an additional sensor component of the mobile robot, wherein the corresponding current additional output from the additional sensor component of the mobile robot is a second type of observation, wherein the second type of observation has a lower fidelity than does the first type of observation.