Robbyant Unveils LingBot-VLA 2.0: A New Frontier in Robotic Embodiment

A Paradigm Shift in Robotic Embodiment

In the rapidly evolving landscape of robotics, the challenge of cross-embodiment manipulation—teaching a single AI model to control diverse robotic hardware—has long been a primary hurdle. Ant Group’s research arm, Robbyant, has taken a significant step toward solving this with the release of LingBot-VLA 2.0. This open-source, 6B parameter vision-language-action (VLA) model is designed to harmonize robotic control, allowing a single intelligence to manage everything from dexterous hands to mobile bases.

By releasing this model under the Apache-2.0 license, Robbyant is signaling a commitment to open science, inviting the global research community to stress-test and build upon their architecture. With its sophisticated training regimen and unique action-space mapping, LingBot-VLA 2.0 is poised to become a foundational tool in the development of general-purpose robotic agents.

The Engineering Behind the 6B Model

At the core of LingBot-VLA 2.0 lies a massive data foundation. The model was pretrained on approximately 60,000 hours of data, a dataset that balances the complexities of physical interaction with the nuanced patterns of human movement. This data is split between 50,000 hours of actual robot trajectories—covering 20 distinct robot configurations—and 10,000 hours of egocentric human video. This hybrid approach allows the model to learn not just the mechanics of movement, but also the intent behind human actions.

One of the most innovative features of this release is the mapping of all embodiments into a single, 55-dimensional canonical action space. Whether the robot is a stationary arm or a complex mobile manipulator, the model interprets commands through this unified lens. This abstraction layer is what enables the "cross-embodiment" capability, effectively decoupling the AI’s reasoning from the specific hardware it controls.

Advanced Architecture: Mixture-of-Experts and Geometric Supervision

To manage the computational load without sacrificing performance, Robbyant has implemented a token-level, auxiliary-loss-free Mixture-of-Experts (MoE) action expert. This architecture allows the model to scale its capacity dynamically. Crucially, the team avoided traditional load-balancing losses, which can often hinder training stability in complex robotic tasks.

Furthermore, the model incorporates dual-query distillation from LingBot-Depth and DINO-Video. This provides the AI with:

• Geometric Supervision: A deep understanding of spatial relationships and 3D depth, essential for precise grasping.
• Temporal Supervision: A future-aware control mechanism that allows the robot to predict the outcomes of its movements, leading to smoother and more intentional navigation.

Performance Benchmarks and Future Outlook

In head-to-head testing against industry standards, LingBot-VLA 2.0 has shown remarkable proficiency. On the GM-100 generalist benchmark, the model outperformed both its predecessor, LingBot-VLA 1.0, and the well-regarded π0.5 model. These results suggest that the combination of massive, diverse data and a unified action space is the correct trajectory for the field.

As robotic systems move away from task-specific programming toward generalized intelligence, models like LingBot-VLA 2.0 will be essential. By standardizing how robots 'understand' their own bodies in relation to the physical world, Robbyant is effectively lowering the barrier to entry for high-level robotics research. Whether in industrial manufacturing, logistics, or household assistance, the ability to deploy a single intelligence across varied hardware configurations represents a milestone in the journey toward true general-purpose robotics.