The robotics company Lingbo Technology, under Ant Group, today announced the full open-source release of its embodied foundation model LingBot-VLA's real-robot post-training toolchain. Development teams can use this toolchain to quickly adapt LingBot-VLA to their own robots and specific tasks using their own data.

 

Currently, the number of open-source models in the field of embodied intelligence continues to grow, but deploying a model on one's own robot still requires a series of adaptation steps. Due to differences in robotic arm configurations, end-effectors, sensor setups, and control interfaces among different robots, development teams often need to carry out extensive engineering work around real-robot deployment. This engineering workflow is usually the core know-how of each team, and has rarely been fully open-sourced in the past.

 

This open-source release targets the core needs in the real-robot adaptation process, covering four key stages: data processing tools that support merging multiple LeRobot datasets and standardize joint dimension mapping, training configurations optimized for real-robot scenarios, offline evaluation tools, and a real-robot deployment module that supports compilation acceleration. The model also provides two versions, with and without depth, allowing development teams to choose based on their own needs.

 

As an embodied base model from Lingbo, LingBot-VLA is pre-trained on 20,000 hours of real-robot data, covering 9 mainstream dual-arm robot configurations, and has cross-body and cross-task generalization capabilities. In both real-robot and simulation evaluations, LingBot-VLA outperforms the industry benchmark π0.5, and has completed multi-machine verification with manufacturers such as Lepu, Songling, and Xinghai Tu.

 

It is reported that LingBot-VLA can achieve high-quality task transfer with just 150 demonstration data points. Thanks to deep optimization of the underlying code library, its training efficiency is 1.5 to 2.8 times that of mainstream frameworks such as StarVLA and OpenPI, further reducing the data and computing cost required for model adaptation.

 

Currently, the LingBot-VLA code repository is open-sourced on GitHub (github.com/Robbyant/lingbot-vla), and the model weights are also released on Hugging Face and ModelScope.