Скачать или смотреть LingBot-VLA: Scaling VLA Models for Robotics

In this AI Research Roundup episode, Alex discusses the paper: 'A Pragmatic VLA Foundation Model' LingBot-VLA is a pragmatic Vision-Language-Action foundation model designed to scale robotic manipulation capabilities using a massive real-world dataset. The model leverages 20,000 hours of training data from nine diverse dual-arm robotic systems to achieve high generalizability. By using a Mixture-of-Transformers architecture, it integrates the Qwen2.5-VL model with a specialized action expert module for seamless control. The framework also employs Flow Matching and vision distillation to enhance spatial awareness and continuous action modeling. Evaluation on the GM-100 benchmark confirms its superior performance across 100 different robotic tasks. Paper URL: https://arxiv.org/abs/2601.18692 #AI #MachineLearning #DeepLearning #Robotics #VLA #FoundationModels #ComputerVision

Resources:
GitHub: https://github.com/robbyant/lingbot-vla
Hugging Face model: https://huggingface.co/robbyant/lingb...
Hugging Face model 2: https://huggingface.co/robbyant/lingb...