A Pragmatic VLA Foundation Model

A new vision-language-action model for robotic manipulation, LingBot-VLA, has been developed using roughly 20,000 hours of real-world data from nine dual-arm robot configurations, according to a paper posted on arXiv [1][2]. The model was trained to generalize across tasks and platforms while maintaining cost efficiency in data and GPU hours required for adaptation [2]. In a systematic assessment on three robotic platforms, each completing 100 tasks with 130 post-training episodes per task, LingBot-VLA outperformed competing approaches [2]. The authors report that the model demonstrates strong performance and broad generalizability, making it suitable for real-world deployment [2]. The accompanying codebase achieves a throughput of 261 samples per second using an eight-GPU training setup, which the researchers state represents a 1.5 to 2.8 times speedup over existing VLA-oriented codebases, depending on the underlying vision-language model [2]. The paper, submitted by Kecheng Zheng and collaborators, was first posted on January 26, 2026, and revised through June 2026 [1]. The work appears on arXiv, an open-access repository that hosts electronic preprints across fields including computer science, physics, and mathematics [7]. As of late 2024, arXiv was receiving approximately 24,000 new submissions per month and had surpassed two million total articles [7]. The platform also supports arXivLabs, a framework launched in 2020 that enables community contributors to build experimental tools integrated directly on article pages [5]. These tools include citation explorers, code finders, and recommender systems, all developed under guidelines that emphasize openness and user data privacy [5][6]. The researchers have released the code, base model, and benchmark data openly, aiming to support more challenging tasks and promote consistent evaluation standards in robot learning [2].