智通财经网·2025-07-25 07:03Core Insights - ByteDance's Seed team launched the GR-3 general robot model, which shows superior operational performance in new environments and object handling compared to the GR-2 model set to release in October 2024 [1][2] - The GR-3 model demonstrates significant improvements in generalization and complex task execution success rates over the industry-leading embodied model π0 [1][4] Model Architecture and Training - The GR-3 model utilizes a MoT+DiT network structure, integrating the "vision-language module" and "action generation module" into a 4 billion parameter end-to-end model, enhancing dynamic instruction following through RMSNorm [2] - The training methodology for GR-3 includes a three-in-one data training approach, utilizing high-quality remote operation data, low-cost human VR trajectory data, and publicly available image-text data to improve generalization capabilities [2] Hardware Development - To maximize the potential of the GR-3 model, ByteDance introduced the ByteMini, a dual-arm mobile robot designed specifically for GR-3, featuring 22 degrees of freedom and a unique wrist ball joint design for enhanced flexibility [3] - The ByteMini includes a multi-camera coordination system for comprehensive situational awareness and a whole-body motion control system to ensure smooth trajectory generation and adaptive force adjustment during tasks [3] Performance Comparison - In comparative testing, GR-3 outperformed π0 in four categories: basic environment, new environment, complex instructions, and new objects, achieving a 17.8% higher success rate in new object handling [4] - GR-3 can elevate the success rate of new object operations from 60% to over 80% with just 10 human trajectory data points, showcasing its high generalization and complex task execution capabilities [4]