This module implements all the data-types and logic required to maintain an intermediary world-graph representation which is used by planner for decision-making. This is a three-tiered tree going from House -> Room -> Furniture, with objects added as leaf nodes which can move around in the tree as an episode evolves.
These are basic building blocks of a world-graph representation, consisting of following node-types:
- Human
- SpotRobot
- Room
- Floor
- Furniture
- objects
We provide two types of graphs for our baseline:
- Privileged world-graph: Objects are found based on observations but the location, segmentation, etc. used to find objects is ground-truth. GT information is also used to reassociate object detections over time. (
world_graph.py) - Non-privileged world-graph: Objects are found using ground-truth segmentation sensor but the location, reassociation over time is done using just RGB and depth sensors without using any privileged sim-level information. (
dynamic_world_graph.py)
graph.py consists of basic graph operations agnostic of entities above. world_graph.py implements all PARTNR specific logic for privileged world-graphs without partial observability, i.e. graph is updated based on full observability each step. dynamic_world_graph.py implements all non-privileged PARTNR logic for maintaining world-graph using partial-observations and non-privileged information source.