The Pitfall of More Powerful Autoencoders in Lidar-Based Navigation
This work highlights a critical pitfall for researchers and practitioners optimizing pretrained autoencoders for lidar-based navigation, showing that better reconstruction does not guarantee better RL performance.
This paper investigates the use of variational autoencoders for generating compact state representations from 2D-lidar scans for reinforcement learning navigation tasks. They propose a novel preprocessing step into a local binary occupancy image, which significantly improves reconstruction capabilities compared to the state of the art. However, they demonstrate that improved reconstruction power does not predictably lead to improved performance in reinforcement learning tasks.
The benefit of pretrained autoencoders for reinforcement learning in comparison to training on raw observations is already known [1]. In this paper, we address the generation of a compact and information-rich state representation. In particular, we train a variational autoencoder for 2D-lidar scans to use its latent state for reinforcement learning of navigation tasks. To achieve high reconstruction power of our autoencoding pipeline, we propose an - in the context of autoencoding 2D-lidar scans - novel preprocessing into a local binary occupancy image. This has no additional requirements, neither self-localization nor robust mapping, and therefore can be applied in any setting and easily transferred from simulation in real-world. In a second stage, we show the usage of the compact state representation generated by our autoencoding pipeline in a simplistic navigation task and expose the pitfall that increased reconstruction power will always lead to an improved performance. We implemented our approach in python using tensorflow. Our datasets are simulated with pybullet as well as recorded using a slamtec rplidar A3. The experiments show the significantly improved reconstruction capabilities of our approach for 2D-lidar scans w.r.t. the state of the art. However, as we demonstrate in the experiments the impact on reinforcement learning in lidar-based navigation tasks is non-predictable when improving the latent state representation generated by an autoencoding pipeline. This is surprising and needs to be taken into account during the process of optimizing a pretrained autoencoder for reinforcement learning tasks.