外观
论文arXiv 2506.082961 个月前
HiBerNAC：用于解耦复杂操作的层级脑启发机器人神经智能体集群
原文链接：HiBerNAC: Hierarchical brain-inspired robotic neural agent collective for disentangling complex manipulation（arXiv 2506.08296 · 发布日期：2026-09-01）
HiBerNAC: Hierarchical brain-inspired robotic neural agent collective for disentangling complex manipulation
HiBerNAC以脑启发层级多智能体结合VLA规划推理，在MuJoCo仿真复杂操作基准验证长程与多路径任务表现。
核心要点
- 复杂操作仍受持续上下文记忆、不确定协同和动态长程规划限制。
- 提出模块化层级脑启发多智能体架构，融合多模态VLA规划推理与反思机制。
- 通过动态智能体专门化形成可扩展集体智能，适应不同任务时长与复杂度。
- MuJoCo仿真中较SOTA VLA平均长程完成时间降23%，多路径成功率为8%–48%。
验证范围与局限：验证范围为基于MuJoCo的仿真基准复杂操作任务，摘要未报告真机实验。
英文摘要 (Abstract)
Recent advances in multimodal vision language action (VLA) models have revolutionized traditional robot learning, enabling systems to interpret vision, language, and action in unified frameworks for complex task planning. However, mastering complex manipulation tasks remains an open challenge, constrained by limitations in persistent contextual memory, multi-agent coordination under uncertainty, and dynamic long-horizon planning across variable sequences. To address this challenge, we draw conceptual inspiration from principles of neural organization and hierarchical decision-making, using them as architectural design metaphors rather than as claims of biologically faithful neural emulation. We present HiBerNAC, a hierarchical brain-inspired robotic neural agent collective that combines: (1) multimodal VLA planning and reasoning with (2) brain-inspired reflection and multi-agent mechanisms, organized as a modular multi-agent architecture conceptually motivated by principles of neural organization, specifically designed for complex robotic manipulation tasks. By leveraging brain-inspired functional modules with decentralized multi-agent collaboration, our approach enables robust and enhanced real-time execution of complex manipulation tasks. In addition, the agentic system has scalable collective intelligence via dynamic agent specialization, adapting its coordination strategy to variable task horizons and complexity. Through extensive experiments (conducted in a MuJoCo-based simulation benchmark) on complex manipulation tasks compared with the state-of-the-art VLA models, we demonstrate that HiBerNAC reduces average long-horizon task completion time by 23%, and achieves non-zero success rates (8%–48%) on multi-path tasks where prior state-of-the-art VLA models consistently failed (differences validated by two-sided Welch’s t -tests, p < 0.05 for correction, OOD, multimodal, and long-horizon categories). These results suggest that principles of neural organization can serve as useful inductive biases for designing robotic manipulation systems. More details are available at our project website .
作者：Heng Zhang, Hongjun Wu, Pengsong Zhang, Jin Wang, Cong Wang
来源
本页事实对应的公开页面，以原文为准。