Recently, Lingbo Technology, a company under Ant Group specializing in embodied intelligence, officially announced the open-source release of its embodied large model LingBot-VLA. The model has demonstrated excellent performance on multiple authoritative benchmark tests and has also released a complete post-training code library, aiming to lower the R&D barriers in the field of embodied intelligence.

Performance: Multi-platform Generalization and Precise Control

LingBot-VLA has shown strong accuracy and generalization capabilities in both real-world environments and simulation scenarios:

Real-world Evaluation: In the GM-100 evaluation benchmark at Shanghai Jiao Tong University, LingBot-VLA achieved an average success rate of 15.7% across three different robot platforms, surpassing Pi0.5's 13.0%.

Spatial Perception Enhancement: After incorporating depth information, its average success rate further increased to 17.3%.

Simulation Leadership: In the RoboTwin2.0 simulation evaluation, facing random disturbances such as lighting and clutter, its operation success rate was 9.92% higher than that of Pi0.5.

Technical Core: Efficient Post-training Toolchain

The advantages of LingBot-VLA are not only reflected in its performance but also in its high training efficiency and migration capability:

Low-cost Migration: Thanks to large-scale pre-training, the model can achieve performance exceeding mainstream models with less data in downstream tasks.

High-throughput Training: The team built an efficient toolchain, which can process 261 samples per second per GPU card under an 8-GPU configuration, achieving a training efficiency 1.5 to 2.8 times that of mainstream frameworks like StarVLA and OpenPI.

Open-source Content: Full Resources Accessible with One Click

Lingbo Technology has made a significant open-source effort, providing full-chain support from weights to tools:

Model Weights: Are now available on Hugging Face and the ModelScope community.

Codebase: The GitHub repository is open, including all code for data processing, efficient fine-tuning, and automated evaluation.

Datasets and Reports: The GM-100 dataset and detailed technical reports are also provided.

The comprehensive open-sourcing of LingBot-VLA offers robot developers a truly practical, efficient, and low-cost VLA model option, and is expected to accelerate the transition of embodied intelligence technology from laboratories to large-scale real-world applications.