Modeling Dynamic Environments with Scene Graph Memory
Andrey Kurenkov, Michael Lingelbach,Tanmay Agarwal, Emily Jin, Chengshu Li, Ruohan Zhang, Li Fei-Fei, Jiajun Wu, Silvio Savarese, Roberto Martín-Martín
Abstract
Embodied AI agents that search for objects in large environments such as households often need to make efficient decisions by predicting object locations based on partial information. We pose this as a new type of link prediction problem: link prediction on partially observable dynamic graphs.
Our graph is a representation of a scene in which rooms and objects are nodes, and their relationships are encoded in the edges; only parts of the changing graph are known to the agent at each timestep. This partial observability poses a challenge to existing link prediction approaches, which we address. We propose a novel state representation -- Scene Graph Memory (SGM) -- with captures the agent’s accumulated set of observations, as well as a neural net architecture called a Node Edge Predictor (NEP) that extracts information from the SGM to search efficiently.
We evaluate our method in the Dynamic House Simulator, a new benchmark that creates diverse dynamic graphs following the semantic patterns typically seen at homes, and show that NEP can be trained to predict the locations of objects in a variety of environments with diverse object movement dynamics, outperforming baselines both in terms of new scene adaptability and overall accuracy.
Our problem setup and proposed method: an agent is queried to find target objects in an unknown dynamic household environment where objects change location over time and objects may be added or removed -- a specific instance of the general problem of link prediction in partially observable dynamic graphs. The agent can observe subsets of the true scene graph, which are aggregated in a Scene Graph Memory which is then passed into the Node Edge Predictor model that produces observation-conditioned posterior probabilities of where the query object is. Finally, the agent uses the posterior probabilities to decide on its next action.
Node Edge Predictor (NEP) model architecture
iGridson, our gridworld implementation of the iGibson 2.0 simulator. An example household scene with four rooms is shown here with various furniture items and objects.
Read the full paper here.
Cite as follows:
@article{kurenkov2023modeling,
title={Modeling Dynamic Environments with Scene Graph Memory},
author={Kurenkov, Andrey and Lingelbach, Michael and Agarwal, Tanmay and Li, Chengshu and Jin, Emily, and Fei-Fei, Li and Wu, Jiajun and Savarese, Silvio, and Mart{\'i}n-Mart{\'i}n, Roberto},
booktitle={International Conference on Machine Learning},
year={2023},
organization={PMLR}
}