From Foundation to Application: Improving VLA Models in Practice

2026-07-07

AI Summary

This work addresses the performance gap of Vision-Language-Action (VLA) foundation models between laboratory settings and real-world environments by introducing LingBot-VLA-2.0. The model leverages a pretraining framework built on 60,000 hours of multi-source data and, for the first time, unifies head, torso, mobile base, and dexterous hand motions into a single whole-body action space. It incorporates predictive dynamics modeling that fuses video semantics with temporal depth-geometry cues. By combining large-scale robot trajectory data and human egocentric videos during pretraining, along with multimodal representation learning and cross-morphology transfer, the approach significantly enhances generalization and long-horizon mobile manipulation performance on the GM-100 benchmark.

Abstract

Despite recent progress of VLA foundation models, the disparity between laboratory conditions and real-world applications continues to impede their practical implementation. To bridge this gap, we present LingBot-VLA 2.0, which advances LingBot-VLA through improvements in three functional domains. (1) Generalization across tasks and embodiments. Compared to the previous version, we revamp the data processing pipeline and curate around 60,000 hours of data for pretraining, including 50,000 hours of robot trajectories spanning 20 robot configurations and 10,000 hours of egocentric human videos. (2) Expanded action space in addition to dual-arm hardware platforms. In particular, our system accommodates degrees of freedom for the heads, waists, mobile bases, and dexterous hands, thereby empowering the robots to tackle more complex tasks in practical scenarios. (3) Predictive dynamics modeling for improved temporal reasoning. Specifically, we formulate future prediction as a proxy task, facilitated by a video representation model for semantic priors and a depth estimation model for geometric cues. Evaluations on the GM-100 benchmark, conducted in a generalist setting, validate the beneficial impact of these proposed modifications. Furthermore, benefiting from the expanded pretraining data that covers whole-body degrees of freedom, LingBot-VLA-2.0 demonstrates strong cross-embodiment long-horizon mobile manipulation capability across the two robotic platforms.

Problem

• VLA models
• real-world application
• embodiment gap
• practical implementation
• generalization

Innovation

• VLA foundation models
• cross-embodiment generalization
• whole-body action space
• predictive dynamics modeling
• egocentric human video pretraining