Modules

We consider CLIER as a modular framework. Therefore, we prepared an idea of the inference pipeline, located in model.inference.InferenceTool. A tool suggests capability of using the following modules:

• Visual Recognition - predicts visual properties of the objects
• Instruction Translation - translates textual instruction to symbolic programs
• Pose Prediction - predicts poses of all the objects in the scene
• Action Planner - given the goal task, returns primitive action to be currently undertaken

Note that for all the modules, a parent class is prepared, and a ground truth loading version can be invoked. In such way ground truth sequences may be generated (an example in inference.py in main directory).

Note also that modules can be adapted or removed, e.g. Action Executor can be removed, and you can use direct control instead of primitive actions.