Motor System

Monty's motor system is contained within the a MotorSystem object. The motor system's job is to generate Action objects to be consumed by a simulator, robot, etc.

The motor system's primary component is a MotorPolicySelector which itself contains one or more MotorPolicy objects. The MotorPolicySelector is responsible for choosing the policy given any information it has available, which includes goals, proprioceptive information, raw observations, and CMP-compliant data. If goals are not provided, the MotorPolicySelector will still select an appropriate default policy. MotorPolicy can be considered the "terminal" object in the sense that they are the only objects within the motor system that produce actions. In contrast, the policy selector mediates action-generation by trafficking data in and out of the policies it chooses.

Currently, we have defined two motor policy selectors:

• SinglePolicySelector: This is the trivial selector. Configured with a single policy, it merely moves data in and out of its sole policy. This is most often used during training (e.g., with the NaiveSpiralScan motor policy) or with the surface agent configurations.
• DistantPolicySelector: This selector has three policies it can choose from. If the selector has a goal that was sent from a learning module, it will choose the JumpToGoal policy which repositions the agent and sensors in the environment so that Monty will be able to observe the target location. If there are no LM-derived goals but there are goals from an SM (e.g. SalienceSM), then this selector will saccade to the target location with the LookAtGoal policy. When there are no goals present, it will default to the InformedPolicyRandomWalk, which merely moves around at random.

For more detailed information on how our policies work, see Policy.

Help Us Make This Page Better

All our docs are open-source. If something is wrong or unclear, submit a PR to fix it!

Make a Contribution

Learn how to contribute to our docs