A comprehensive survey in the journal Robot examines how large language models (LLMs) and multimodal large models are reshaping intelligent robotics across perception, navigation, and manipulation. The authors, from institutions including Huazhong University of Science and Technology and Concordia University, systematically review recent advances and outline a roadmap for building general-purpose, cognition-enhanced robotic systems.
Key takeaways
- Perception: Multimodal fusion and language-spatial joint reasoning enable deeper understanding of environmental semantics and geometry, improving object recognition and scene interpretation in unstructured settings.
- Navigation: Chain-of-thought task decomposition and commonsense reasoning allow robots to parse ambiguous natural-language instructions and autonomously explore unknown environments, moving beyond pre-programmed path planning.
- Manipulation: Vision-language-action (VLA) models coupled with physical commonsense enhance dexterity and adaptability in complex interactive tasks, supporting dynamic strategy adjustment based on real-time visual feedback.
- Paradigm shift: Large models drive robotics from “perception-driven” to “cognition-driven” operation, significantly improving contextual reasoning and autonomous decision-making.
- Core challenges: Cross-modal alignment accuracy, real-time performance, safety and reliability, and simulation-to-real (Sim2Real) generalization remain unresolved, limiting widespread industrial deployment.
For manufacturers and system integrators, the survey underscores the potential of LLM-integrated robots to handle more ambiguous tasks and dynamic environments, but also cautions that practical adoption requires overcoming robustness and real-time constraints. The paper provides a structured technical reference for evaluating emerging embodied AI solutions.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-07-15 · “大模型与智能机器人的融合：智能感知、导航与操作”