Learning-Augmented Model-Based Planning
for Visual Exploration


Yimeng Li*
Arnab Debnath*
Gregory Stein
Jana Kosecka
George Mason University

IROS 2023

[Paper]
[Video]
[Code]


At each step, the agent will receive a 360 panorama or 90 degree egocentric observation. Here in the video, we show a running example of our model exploring a novel scene using a frontier-based approach. A frontier is a boundary between free and unknown space, as denoted by the green pixels in the video. Yellow pixels are the selected frontier.


We consider the problem of time-limited robotic exploration in previously unseen environments where explo- ration is limited by a predefined amount of time. We propose a novel exploration approach using learning-augmented model- based planning. We generate a set of subgoals associated with frontiers on the current map and derive a Bellman Equation for exploration with these subgoals. Visual sensing and advances in semantic mapping of indoor scenes are exploited for training a deep convolutional neural network to estimate properties associated with each frontier: the expected unobserved area beyond the frontier and the expected timesteps (discretized actions) required to explore it. The proposed model-based planner is guaranteed to explore the whole scene if time permits. We thoroughly evaluate our approach on a large-scale pseudo-realistic indoor dataset (Matterport3D) with the Habitat simulator. We compare our approach with classical and more recent RL-based exploration methods, demonstrating its clear advantages in several settings.

Overview

Learning-Augmented Model-Based Frontier Exploration (LFE)

This diagram gives an overview of our learning-augmented model-based exploration algorithm, which allows us to compute the expected value of each action in a computationally feasible way for planning through an unknown environment. We use learning modules to estimate the terms R_A, D_in and D_out, thereby introducing prior knowledge about environment regularities into the decision-making procedure.

Architecture of the entire exploration system

The detected and the selected frontiers are drawn in green and yellow respectively. Our proposed Learning-augmented model-based frontier-based exploration is used for the frontier selection module. We design two learning modules to estimate the frontier properties. The UNet model takes in the local occupancy map and local semantic map and estimates values for all the visible frontiers on the map. The ResNet-18 model takes in the egocentric depth and semantic observation corresponding to a frontier and estimates the values only for this frontier.

Compare different planning strategies

In frontier-based methods, the robot’s actions correspond to navigation to frontiers. These actions rely on the partial map to plan and reliably navigate through space the robot has already seen. Existing Approaches greedily select the next frontier-action. Classical non-learning greedy exploration from Brian greedily select the next frontier-action via simple heuristics by navigating to the nearest frontier. Recent work called PONI uses visual sensing and replaces simple heuristics with a learning-based module.

Learning Module

We develop a learning module for estimating frontier properties from a partial map or egocentric views.
For the map based model, We use a U-net architecture where the input to the U-Net model is the currently observed occupancy and semantic map. For the view based model, the input to the ResNet-18 model is the egocentric depth and semantic observation.

Perception Module

Results: compare LFE with the baseline Greedy approach

Citation

                        @inproceedings{li2023learning,
                          title={Learning-augmented model-based planning for visual exploration},
                          author={Li, Yimeng and Debnath, Arnab and Stein, Gregory J and Ko{\v{s}}eck{\'a}, Jana},
                          booktitle={2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
                          pages={5165--5171},
                          year={2023},
                          organization={IEEE}
                        }

Acknowledgements

We thank members of the GMU Vision and Robotics Lab and RAIL.
This webpage template was borrowed from some colorful folks.