On July 8th, Ant Lingbo Technology announced the upgrade and open sourcing of the new generation embodied base model LingBot VLA 2.0. As a comprehensive upgrade to the open-source version LingBot VLA 1.0 released in January this year, LingBot VLA 2.0 incorporates 60000 hours of high-quality real physics data during the pre training phase, covering 20 robot configurations from 17 mainstream robot brands, and expanding support for degrees of freedom such as head, waist, end effector, and moving base. Significant improvements have been achieved in configuration generalization, degree of freedom support, and landing efficiency.

The current embodied intelligence industry is advancing rapidly, with the "cerebellum" and hardware ontology accelerating their evolution. However, the industry's "universal brain" remains the core constraint for the landing of large-scale industries. From the perspective of the 'brain', breakthroughs are urgently needed in both model capability and the efficiency and cost of model implementation.

According to the technical report, LingBot VLA 2.0 supports robot brands such as Leju, Zhiyuan, Yushu, Songling, Xinghaitu, Galaxy General, Xingchen, and Ruilman during the pre training phase Franka、 17 robot manufacturers including Ark, Beijing Humanoid, Fourier, Magic Atom, Chihiro, Zero Power, Feixi, Qinglong, etc., covering various forms such as single arm/double arm, bipedal/wheeled.

In terms of degree of freedom support, LingBot VLA 2.0 comprehensively expands support for degrees of freedom such as head, waist, end effector (hand), and mobile chassis.

In terms of dual arm operation, based on the GM-100 evaluation from Shanghai Jiao Tong University, LingBot VLA 2.0 outperforms π 0.5 and GR00T N1.7 in overall average task progress score and success rate on two dual arm robot platforms, AgileX Cobot Magic and Galaxea R1 Pro. In this evaluation, all participating models were deployed as a single generalized model, without any specialized adjustments for specific tasks. This result demonstrates the stronger dual arm collaborative operation ability and cross ontology, multi task generalization ability of LingBot VLA 2.0.

(Image caption: LingBot VLA 2.0 leads in overall performance in GM-100 evaluation)

In terms of mobility, LingBot VLA 2.0 is based on two configurations - Ark robotic arm+Songling chassis, and Stardust intelligent Astribot S1, and has undergone preliminary comparative testing with π 0.5. The results show that LingBot-VLA 2.The task progress score and success rate of 0 in long-range mobile operation tasks are leading, especially in more challenging cross domain scenarios, demonstrating stronger ability to advance long sequence tasks and generalize mobile operations.

In mobile operation evaluation, tasks are broken down into multiple consecutive sub steps, with each step assigned different scores based on difficulty and importance. Robots can obtain corresponding scores by completing the corresponding steps, and the final total score reflects their task advancement ability in long sequence tasks. Compared to simply calculating the final success rate, this scoring method can more finely measure the comprehensive ability of the model in different aspects such as movement, dual arm cooperation, grasping, placement, door opening, and cleaning.

(Image caption: LingBot VLA 2.0 has significant advantages in cross domain scenarios when it comes to long-range mobile operation tasks.)

Supporting these capability upgrades are larger scale, higher quality data systems and better training architectures: Ant Lingbo cleanses 50000 hours of high-quality real machine data from 90000 hours of data, and extracts 10000 hours of effective data from 20000 hours of first person human operation data, making the total amount of pre training data reach 60000 hours.

At present, the industry has gradually entered the pilot stage of industrial landing, and efficient post training has become a key constraint factor for landing. The LingBot VLA 2.0 synchronized open source version is more efficient for training, and the inference time is controlled within 130 milliseconds on RTX 4090.

It is reported that Ant Lingbo has partnered with ecological partners such as Leju and Titanium Tiger, as well as ecological customer partners such as Guoda Pharmacy and Longsheng, to conduct comprehensive commercial landing tests in retail sorting, logistics sorting, industrial and other landing scenarios. On the other hand, Ant Lingbo collaborates with data alliance ecological partners such as Jianzhi Technology to jointly build a standardized data system. A embodied intelligence ecosystem is taking shape, with cross configuration VLA base models as its core and deep involvement of ontology vendors and data institutions.

Currently, LingBot VLA 2.0 is open source. Developers can obtain model weights on Hugging Face and the Magic Community, and download open source code on GitHub. It is reported that Ant Lingbo will also launch a series of developer activities in the next step, and simultaneously release technology kits that are more suitable for developers.