LingBot-VLA trained on 60,000 hours of physical data
Robbyant, an embodied AI company within Ant Group, announced in Shanghai on 8th July 2026 that it had upgraded and open-sourced LingBot-VLA 2.0. Earlier, the company released LingBot-VLA 1.0 in January 2026.
LingBot-VLA 2.0 was pre-trained on 60,000 hours of high-quality real-world physical data. That total includes 50,000 hours of cleaned real-robot interaction data and 10,000 hours of distilled first-person human manipulation data.
Training data came from 20 robot morphologies across 17 manufacturers. Those manufacturers include Leju, AgiBot, Unitree, AgileX, Galaxea, Galbot, Astribot, RealMan, Franka, ARX, X-Humanoid, Fourier, MagicLab, Spirit AI, Zerith, Flexiv and Qinglong.
Across that dataset, Robbyant covered single-arm, dual-arm, bipedal and wheeled robots. The aim is a more scalable “universal brain” for real-world robotics.
LingBot-VLA 2.0 also expands degrees of freedom support. It covers the head, waist, end-effectors such as hands, and the mobile chassis for whole-body control.
Robbyant tied the release to a broader deployment problem. While robot hardware and control systems have advanced fast, the company identified a universal robotics brain as a bottleneck for industrial-scale use.
Shanghai Jiao Tong GM-100
On Shanghai Jiao Tong University’s GM-100 benchmark, LingBot-VLA 2.0 recorded leading average task progress scores and success rates in dual-arm manipulation. Tests used the AgileX Cobot Magic and Galaxea R1 Pro platforms.
In those GM-100 runs, LingBot-VLA 2.0 outperformed π0.5 and GR00T N1.7. Robbyant presented that result as evidence of stronger cross-morphology generalization.
Separate long-horizon mobile manipulation tests also showed gains. On the ARX Arm plus AgileX Chassis system and the Astribot S1 platform, LingBot-VLA 2.0 beat π0.5 in both task progress scores and success rates.
Those cross-domain tests point to stronger long-sequence task execution and broader mobile manipulation generalization. Robbyant also released a version tuned for more efficient post-training and deployment.
Latency stays under 130 milliseconds on an RTX 4090. As a result, Robbyant said the model lowers the cost barrier for commercial deployment.
Robbyant is already testing business uses for LingBot-VLA 2.0. Current work includes hardware collaborations with Leju and Ti5 Robot, plus enterprise projects with GuoDa Drugstore and Longsheng Technology.
Meanwhile, the company is also working with GenRobot.ai on standardised data ecosystems. Those commercial pilots cover retail sorting, logistics and industrial automation.
