Adaptive entropy soft actor-critic with multi-sensor fusion for autonomous robot navigation in dynamic environments
Nature Journal of Emerging Sciences Technologies and Innovations · 30.07.2026 · 9 (5) · 517-550
DOI: 10.65752/bkrc9056 ↗Kurz gefasst
Fachartikel in der Zeitschrift Nature Journal of Emerging Sciences Technologies and Innovations, online erschienen am 30.07.2026. Verfasst von Cecilia Augustine Abi und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Autonomous robots in a dynamic environment populated by humans have to manage the exploration-exploitation tradeoff, fuse heterogeneous and noisy sensor measurements, steer clear of obstacles, and transfer policies trained in a simulator onto a physical hardware robot. Existing Soft Actor-Critic (SAC) algorithm employs a fixed entropy regularization coefficient, incapable of handling a changing policy distribution. This may severely affect learning efficiency, as well as robustness to the environment or safe navigation to the robot when deployed physically. We propose a framework called Adaptive Entropy Soft Actor-Critic (Adaptive SAC) that adapts the entropy coefficient during the learning process based on the actual policy entropy and fuses sensor measurements from wheel-encoders, an IMU, an ultrasonic sensor, infrared sensor, and cameras into a single representation of the navigation state. The main contributions are as follows: 1) an online update rule for the entropy coefficient, with carefully defined update boundaries and parameters; 2) a state representation module designed to combine information from five heterogeneous sensors (five streams of sensor data, each synchronized and fused at the feature level); 3) a composite reward function consisting of elements for goal achievement, safety, collision prevention, and comfort toward humans. The framework is evaluated in simulation (using Webots) over different static, dynamic, human presence, and previously unseen environments, and is also tested on a real 4WD robot and compared to Standard SAC and Safety SAC baselines. The Adaptive SAC reaches convergence in 290 episodes (410 and 400 episodes for Standard SAC and Safety SAC respectively) and achieves a 7% simulated collision rate, 92% success in generalization to unknown environments, and an average human aware distance of 1.65 m. A physical implementation of the framework reached 84% of task completion, a small decrease in performance from 91% in simulation and a 92.31% transfer retention rate, which suggests a partial, though usable, sim-to-real transfer. The results, supported by ANOVA tests (p < 0.001), show significant improvement over standard and safety aware reinforcement learning agents in terms of efficiency, safety and generalization for robot navigation problems.
Zitieren
APA
Abi, C. A., Essien, J., & Amaku, A. E. (2026). Adaptive entropy soft actor-critic with multi-sensor fusion for autonomous robot navigation in dynamic environments. Nature Journal of Emerging Sciences Technologies and Innovations, 9(5), 517-550. https://doi.org/10.65752/bkrc9056
Vancouver
Abi CA, Essien J, Amaku AE. Adaptive entropy soft actor-critic with multi-sensor fusion for autonomous robot navigation in dynamic environments. Nature Journal of Emerging Sciences Technologies and Innovations. 2026;9(5):517-550. doi:10.65752/bkrc9056
BibTeX
@article{Abi2026,
author = {Abi, Cecilia Augustine and Essien, Joe and Amaku, Amaku Effiong},
title = {Adaptive entropy soft actor-critic with multi-sensor fusion for autonomous robot navigation in dynamic environments},
journal = {Nature Journal of Emerging Sciences Technologies and Innovations},
year = {2026},
volume = {9},
number = {5},
pages = {517--550},
doi = {10.65752/bkrc9056}
}
EUVIMED