Summary

LingBot-VLA 2.0 is a Vision-Language-Action foundation model advancing from pre-training to practical robot applications. It targets researchers and engineers seeking enhanced generalization across diverse robot embodiments and tasks, an expanded action space, and improved predictive dynamics for reliable real-world deployment. The model offers a robust framework for developing sophisticated robotic agents.

How It Works

The core approach leverages a redesigned data pipeline curating approximately 60,000 hours of robot trajectories and egocentric videos. LingBot-VLA 2.0 unifies heterogeneous embodiments into a 55-dimensional canonical state/action vector, supporting a wide range of robotic hardware. Predictive dynamics modeling uses DINO-Video (semantic temporal priors) and LingBot-Depth (geometric cues) for future state prediction. Sparse Mixture-of-Experts (MoE) layers within the action expert enhance cross-embodiment scaling. Dual-query distillation from LingBot-Depth and DINO-Video refines perceptual understanding by incorporating current and future scene queries.

Quick Start & Requirements

• Requirements: Miniconda/Anaconda, Python 3.12, PyTorch 2.8.0. Conda must be initialized.
• Installation: Clone the repository and run bash tools/create_train_env.sh.
• Pre-trained Weights: Available for LingBot-VLA 2.0 (6B).
• Additional Weights for Training: Qwen3-VL-4B-Instruct, MoGe-2-vitb-normal, LingBot-Depth, DINO-VIDEO teacher checkpoints are required.
• Documentation: Technical report and pre-trained weights are prepared.

Highlighted Details

• Performance: Achieves 66.2 / 34.4 on GM-100 Bimanual Manipulation, surpassing previous versions. Demonstrates strong out-of-domain performance on long-horizon mobile manipulation tasks.
• Unified Action Space: Maps diverse robotic hardware (arms, hands, mobile bases) to a 55-dimensional canonical state/action vector.
• MoE Action Expert: Improves scalability and generalization across embodiments.
• Dual-Query Distillation: Enhances causal inference by distilling current and future perceptual information.

Maintenance & Community

No specific community channels or maintenance details are provided in the README.

Licensing & Compatibility

• License: Apache-2.0 License.
• Compatibility: The permissive Apache-2.0 license generally allows for commercial use and integration into closed-source projects.

Limitations & Caveats

Presented as a foundation model transitioning towards real-world applications, suggesting an active research and development phase. Specific limitations, unsupported platforms, or known bugs are not detailed.