A Pragmatic VLA Foundation Model

TL;DR

LingBot-VLA tackles the challenge of scalable, generalizable robotic manipulation by pre-training a Vision-Language-Action foundation model on about $20{,}000$ hours of real-world data from $9$ dual-arm embodiments and evaluating across $3$ platforms on $GM{-}100$-style benchmarks. It combines a pre-trained VLM with a learned action expert in a Mixture-of-Transformers, trained via Flow Matching and enhanced by depth-guided distillation, while applying advanced distributed training techniques (FSDP with shard groups) and operator-level optimizations to achieve high throughput on multi-GPU clusters ($ ext{throughput} = $ $261$ samples/s per GPU on an $8$-GPU setup). The results show state-of-the-art performance and strong cross-embodiment generalization, with notable gains from incorporating depth information in both real-world and simulation benchmarks. By releasing code, base models, and benchmarks, the work aims to accelerate real-world deployment and foster rigorous evaluation standards in large-scale VLA robotics research.

Abstract

Offering great potential in robotic manipulation, a capable Vision-Language-Action (VLA) foundation model is expected to faithfully generalize across tasks and platforms while ensuring cost efficiency (e.g., data and GPU hours required for adaptation). To this end, we develop LingBot-VLA with around 20,000 hours of real-world data from 9 popular dual-arm robot configurations. Through a systematic assessment on 3 robotic platforms, each completing 100 tasks with 130 post-training episodes per task, our model achieves clear superiority over competitors, showcasing its strong performance and broad generalizability. We have also built an efficient codebase, which delivers a throughput of 261 samples per second per GPU with an 8-GPU training setup, representing a 1.5~2.8$\times$ (depending on the relied VLM base model) speedup over existing VLA-oriented codebases. The above features ensure that our model is well-suited for real-world deployment. To advance the field of robot learning, we provide open access to the code, base model, and benchmark data, with a focus on enabling more challenging tasks and promoting sound evaluation standards.

Figures (9)

• Figure 1: Overview of LingBot-VLA. We scale dual-arm robot data collected in the real world for pre-training. LingBot-VLA can be easily and efficiently transferred to downstream tasks. Moreover, we conduct a systematic assessment across three robotic embodiments, which demonstrates the clear superiority of our model.
• Figure 2: Visualization of pre-training dataset used by LingBot-VLA.
• Figure 3:
• Figure 4:
• Figure 6: Qwen2.5-VL-3B-$\pi$ model