Building Generalizable Agents with a Realistic and Rich 3D Environment (1801.02209v2)

Published 7 Jan 2018 in cs.LG and cs.AI

Abstract: Teaching an agent to navigate in an unseen 3D environment is a challenging task, even in the event of simulated environments. To generalize to unseen environments, an agent needs to be robust to low-level variations (e.g. color, texture, object changes), and also high-level variations (e.g. layout changes of the environment). To improve overall generalization, all types of variations in the environment have to be taken under consideration via different level of data augmentation steps. To this end, we propose House3D, a rich, extensible and efficient environment that contains 45,622 human-designed 3D scenes of visually realistic houses, ranging from single-room studios to multi-storied houses, equipped with a diverse set of fully labeled 3D objects, textures and scene layouts, based on the SUNCG dataset (Song et.al.). The diversity in House3D opens the door towards scene-level augmentation, while the label-rich nature of House3D enables us to inject pixel- & task-level augmentations such as domain randomization (Toubin et. al.) and multi-task training. Using a subset of houses in House3D, we show that reinforcement learning agents trained with an enhancement of different levels of augmentations perform much better in unseen environments than our baselines with raw RGB input by over 8% in terms of navigation success rate. House3D is publicly available at http://github.com/facebookresearch/House3D.

View on arXiv

Authors (4)

Yi Wu (171 papers)
Yuxin Wu (30 papers)
Georgia Gkioxari (39 papers)
Yuandong Tian (128 papers)

Citations (334)

View on Semantic Scholar

Summary

Overview of "Building Generalizable Agents with a Realistic and Rich 3D Environment"

This paper introduces House3D, a comprehensive virtual environment designed to improve the generalization capabilities of reinforcement learning (RL) agents in 3D navigation tasks. House3D consists of over 45,000 diverse indoor scenes derived from the SUNCG dataset, offering a wide array of realistic house layouts and fully labeled 3D objects. The platform aims to address the challenge of training RL agents that need to generalize beyond the specific environments they are trained in.

Key Contributions

The primary contribution of the paper is the development of House3D, an extensible and rich 3D environment. House3D facilitates various forms of data augmentation, enabling agents to become more robust to a range of environmental variations. The authors leverage the platform to introduce a navigation task termed RoomNav, where agents must navigate to a specified room type based on high-level semantic instructions.

Experimental Framework

The experimentation within the paper focuses on the RoomNav task, where agents, starting randomly positioned in a virtual house, are tasked with reaching a specified room using semantic cues. The agents are trained using gated-CNN and gated-LSTM policies employing advanced RL techniques such as A3C and DDPG. These models incorporate gated attention mechanisms to effectively process both visual inputs and semantic instructions.

In testing the effectiveness of House3D, a subset of houses was used for training, and models were evaluated based on their success rates in unseen environments. Notably, the gated-LSTM model achieved a 35.8% success rate in navigation tasks within 50 unseen environments, outperforming the baseline approach by 10%.

Augmentation Techniques

The authors explore several augmentation techniques:

Pixel-Level Augmentation: Domain randomization is employed to alter the colors and textures within virtual environments, enhancing robustness against visual variation.
Task-Level Augmentation: Multi-target learning is used to train agents on a variety of auxiliary tasks, leading to improved generalization.
Scene-Level Augmentation: Training on an increased number of diverse environments mitigates overfitting and bolsters the agents' performance in novel scenes.

Implications and Future Directions

The emergence of House3D offers significant implications for the design and training of RL agents, especially in the context of applications requiring robust navigation capabilities. By maintaining the availability of thousands of varied environments, House3D not only aids in evaluating the semantic-level generalization of agents but also paves the way for further exploration in real-world robotic tasks.

Future developments may focus on enhancing the fidelity and interaction complexity within House3D, potentially integrating real-world datasets and advancing toward seamless simulation-to-reality transfer. The continued refinement of augmentation techniques, coupled with the exploration of hierarchical task policies, could further enhance agent generalization, fostering advancements in autonomous systems across various domains.

PDF Markdown

Related Papers

Find Related Papers

GitHub

GitHub - facebookresearch/House3D: a Realistic and Rich 3D Environment (1,200 stars)