LingBot-VLA 2.0: A 6B model bridging robot labs and real-world action
Ant Group’s Robbyant just open-sourced LingBot-VLA 2.0, a 6-billion-parameter Vision-Language-Action (VLA) model designed to make robot policies work beyond the lab. The release bundles a technical report, Apache-2.0 code, and a “native depth” checkpoint that turns camera images plus a text instruction into robot actions—all in about 130 milliseconds on a GeForce RTX 4090D.
From lab to factory floor
Most VLA models falter when they leave controlled environments. LingBot-VLA 2.0 tackles that gap with three upgrades: broader generalization, a richer action space, and predictive dynamics modeling. The team trained on roughly 60,000 hours of curated data—50,000 hours of robot trajectories plus 10,000 hours of egocentric human videos—spanning 20 different robot configurations from single-arm rigs to full humanoids.
A unified skeleton for many bodies
Robots come with different joints and kinematics. LingBot-VLA 2.0 smooths this out with a fixed, 55-dimensional canonical state/action vector that covers arms, grippers, waists, heads, mobile bases, and even mobility signals. Missing limbs are simply padded, letting the same model drive diverse hardware without retraining.
Sparse experts, sharp compute
At inference, the model uses a sparse Mixture-of-Experts action head inspired by DeepSeek-V3. Only the top-K experts activate per token, keeping compute predictable while correcting load imbalance with per-expert bias terms. The backbone is Qwen3-VL-4B-Instruct, distilled from two teacher models—LingBot-Depth and DINO-Video—during training.
Explore the model on Hugging Face → Read the technical report →
Why it matters
LingBot-VLA 2.0 signals a shift from brittle research demos to deployable robot policies. By unifying diverse hardware under one action vocabulary and using sparse MoE to tame compute, Robbyant makes it easier for labs and factories to swap in new robots without rewriting the policy. The real stakes are lower integration costs and faster iteration—key levers for scaling robotics beyond today’s constrained setups.
Source: MarkTechPost. AI-assisted editorial synthesis — TechnoExpress.