银河通用&清华推出DexNDM，用神经动力学重塑灵巧操作

Core Insights - The article discusses the development of DexNDM, a new method aimed at solving the sim-to-real challenge in dexterous robotic manipulation, particularly in achieving stable in-hand rotation of various objects [2][5][31] Group 1: Background and Challenges - High dexterity in remote operation of complex tools, such as using a screwdriver or hammer, has been a long-standing challenge in robotics [4] - Traditional direct mapping remote operation methods are limited to simple tasks and cannot handle complex manipulations requiring fine motor skills [4] Group 2: DexNDM Methodology - DexNDM proposes a semi-autonomous remote operation paradigm that breaks down complex tasks into stable, reliable atomic skills that robots can execute autonomously [5] - The method focuses on learning general, stable atomic skills for in-hand object rotation, covering a wide range of scenarios including challenging elongated and small objects [5][14] Group 3: Key Features and Achievements - DexNDM achieves unprecedented dexterity by enabling continuous rotation of elongated objects and intricate manipulation of small objects under challenging wrist postures [7][14] - The method demonstrates superior performance in manipulating complex geometries compared to previous works, even with more general hardware [14] - It showcases high adaptability to various wrist postures and rotation axes, allowing for precise control regardless of the mechanical hand's orientation [17] Group 4: Robustness and Practical Applications - The DexNDM system exhibits high dexterity and robustness, successfully performing complex tool usage tasks such as tightening screws and assembling furniture [21] - The system's robustness allows it to handle long-horizon assembly tasks without interruption, even in the presence of unforeseen scenarios [21] Group 5: Innovations in Data Collection and Modeling - DexNDM employs a joint-wise neural dynamics model that effectively fits real-world data to bridge the gap between simulation and reality [24] - An automated data collection strategy, termed "chaos box," is utilized to gather diverse interaction data with minimal human intervention [28] - The training of a residual policy network is implemented to compensate for the dynamics gap between simulation and real-world applications [30]