Ant Lingbo Open-Sources LingBot-VLA2.0 for 20 Robot Types

新智元 · wechat · 2026-07-08

Ant Group's embodied intelligence subsidiary Ant Lingbo has open-sourced the embodied manipulation foundation model LingBot-VLA 2.0, directly tackling the cross-embodiment generalization challenge of "one brain for many machines." Version 2.0's pre-training covers over 20 robot configurations across 17 mainstream domestic and international brands (including Unitree, AgiBot, Leju, Stardust, etc.), expanding real-world data from 20,000 hours in 1.0 to 50,000 hours.

The team highlighted the current robotics issue of "overheated bodies, missing brains": achievements like backflips and marathons are mostly localized optimizations in closed environments, requiring costly re-adaptation from scratch for each new embodiment. The LingBot-VLA series is positioned as an "industry-universal brain." Version 1.0 achieved a pre-training throughput of 261 samples/second per card using 8-card FSDP parallelism, accelerating 1.5-2.8 times faster than mainstream VLA codebases and drastically reducing post-training costs.

Related event: Ant Lingbo Open-Sources LingBot-VLA 2.0 for Multi-Robot Generalization(18 posts)→

Original post →

More from Embodied

Embodied channel →