机器人学：2025–2026 年度综述

An annual survey of the cs.RO arXiv category, covering research published from September 2025 – August 2026. Built on a close-reading set of 120 papers and organised into 44 chapters, it maps the year by the field's research sub-directions.

Book framework (share of the year's output):

• End-to-End Autonomous Control and Vision-Language-Action Models (19.5%) — Explores unified end-to-end policies, diffusion models, and vision-language-action frameworks for efficient autonomous manipulation and driving.
• Safe and Robust Model Predictive Control (6.1%) — Develops robust, adaptive, and safety-filtered model predictive control techniques to handle dynamic uncertainties and constrained trajectories.
• Real-Time Perception, Mapping, and Neural Scene Reconstruction (10.9%) — Addresses real-time object detection, semantic mapping, and neural scene reconstruction using LiDAR and graph-based methods for generalizable environments.
• Dexterous Manipulation and Humanoid Whole-Body Control (8.6%) — Investigates tactile sensing, dexterous grasping, and whole-body control strategies for humanoid and soft robots in complex physical interactions.
• Hierarchical Task Planning and Embodied Reasoning (4.2%) — Leverages large language models and hierarchical reasoning structures for long-horizon task planning and multi-agent navigation.
• Visual SLAM and Robust Pose Estimation (3.9%) — Enhances visual SLAM and state estimation using Gaussian splatting, neural filters, and robust pose tracking under motion blur and occlusion.
• Deep Reinforcement Learning for Motion Control (3.3%) — Applies deep reinforcement learning, including model-based and offline methods, to optimize low-level motion control and multi-agent coordination.
• Zero-Shot Vision-Language Navigation (0.7%) — Focuses on zero-shot transfer and reasoning in vision-and-language navigation tasks using imaginative frameworks and neurosymbolic planning.
• Active Exploration and Adaptive Sensing Strategies (0.3%) — Investigates active sensing and exploration algorithms to minimize uncertainty and optimize information gain in space and disaster response scenarios.
• Human-Robot Interaction and Multimodal Skill Learning (41.5%) — Focuses on developing multimodal datasets, digital twins, and natural language interfaces to facilitate seamless human-robot interaction and skill adaptation.

Each sub-direction is covered across three chapters — how it framed its questions, the strength of its key findings, and its open debates and frontiers — so the reader gets a traceable picture of what this part of the field actually produced over the year.