SNeRL: Semantic-aware Neural Radiance Fields for Reinforcement Learning
This work addresses the problem of improving reinforcement learning performance for agents by incorporating human-intuitive semantic understanding into 3D representations, representing an incremental advancement over existing methods.
The paper tackles the problem of sub-optimal performance in reinforcement learning due to ineffective 3D environment representations by introducing SNeRL, which jointly optimizes semantic-aware neural radiance fields with a convolutional encoder to learn 3D-aware neural implicit representations from multi-view images, outperforming previous pixel-based and recent 3D-aware representations in both model-free and model-based reinforcement learning.
As previous representations for reinforcement learning cannot effectively incorporate a human-intuitive understanding of the 3D environment, they usually suffer from sub-optimal performances. In this paper, we present Semantic-aware Neural Radiance Fields for Reinforcement Learning (SNeRL), which jointly optimizes semantic-aware neural radiance fields (NeRF) with a convolutional encoder to learn 3D-aware neural implicit representation from multi-view images. We introduce 3D semantic and distilled feature fields in parallel to the RGB radiance fields in NeRF to learn semantic and object-centric representation for reinforcement learning. SNeRL outperforms not only previous pixel-based representations but also recent 3D-aware representations both in model-free and model-based reinforcement learning.