Research on Robust Reinforcement Learning Methods for Embodied Perception of Robotic Arms in Unstructured Environments
Journal of Engineering System · 2026-09 · 4 (3) · 77-86
DOI: 10.62517/jes.202602307 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Engineering System, 2026-09. Verfasst von Zhiyao Wang. Die Arbeit nennt 12 Literaturangaben.
Abstract (Originalsprache)
Embodied perception is a core capability for robotic arms to operate autonomously in unstructured environments. However, traditional methods struggle to effectively distinguish between lighting changes, sensor noise accumulation, and object occlusions due to complex disturbances, potentially leading to operational errors or task failure. To address these challenges, this paper proposes a robust optimization framework for embodied perception in robotic arms based on deep reinforcement learning, systematically evaluating the contributions of each module through ablation studies. The problem of perceptual robustness is modeled as a Markov Decision Process (MDP), with a multimodal attention fusion module designed to incorporate signal-to-noise ratio-aware perception, and contrastive learning employed to enhance feature discriminability. Perceptual confidence is introduced as a state variable, enabling a dynamic perception gain modulation mechanism that achieves end-to-end co-optimization of perception weights and action policies. Additionally, an intrinsic motivation exploration strategy based on prediction error is integrated to alleviate the challenge of sparse rewards in unstructured scenarios. A large-scale parallel training environment is built on the NVIDIA IsaacGym platform, using the UR5e robotic arm to generate diverse task scenarios involving extreme lighting conditions, texture variations, and random occlusions for grasping, placing, and assembly tasks. Ablation results demonstrate that the dynamic perception gain mechanism significantly improves success rates under both normal and strong interference conditions; the intrinsic motivation exploration strategy effectively enhances exploration efficiency in sparse-reward settings; and the multimodal attention fusion module provides a solid foundation for gain adjustment by estimating weight distributions. The complete framework outperforms standard baselines such as DDPG across all three task categories, maintaining low perceptual error levels even under extreme occlusion conditions.
Zitieren
APA
Wang, Z. (2026). Research on Robust Reinforcement Learning Methods for Embodied Perception of Robotic Arms in Unstructured Environments. Journal of Engineering System, 4(3), 77-86. https://doi.org/10.62517/jes.202602307
Vancouver
Wang Z. Research on Robust Reinforcement Learning Methods for Embodied Perception of Robotic Arms in Unstructured Environments. Journal of Engineering System. 2026;4(3):77-86. doi:10.62517/jes.202602307
BibTeX
@article{Wang2026,
author = {Wang, Zhiyao},
title = {Research on Robust Reinforcement Learning Methods for Embodied Perception of Robotic Arms in Unstructured Environments},
journal = {Journal of Engineering System},
year = {2026},
volume = {4},
number = {3},
pages = {77--86},
doi = {10.62517/jes.202602307}
}
EUVIMED