How LingBot‑VLA 2.0 Powers 20 Robot Configurations with an Open‑Source Embodied Brain

After a five‑month hiatus, Ant LingBot unveils LingBot‑VLA 2.0, a next‑generation embodied VLA model that targets the industry‑wide challenge of running a single brain on many robot bodies. The model supports 20 distinct robot configurations—including single‑arm, dual‑arm, wheeled, and biped platforms—by extending action coverage to head, waist, end‑effector and mobile chassis degrees of freedom.

Architecture upgrades : LingBot‑VLA 2.0 adopts a token‑level loss‑free Mixture‑of‑Experts (MoE) to separate shared action logic from robot‑specific dynamics, using token‑level sigmoid routing so each token can activate multiple experts. It also adds a dual‑query distillation framework: one query learns spatial cues from the in‑house LingBot‑Depth depth model, and the other learns temporal and future‑state cues from DINO‑Video, a teacher built on Meta’s DINOv3 with causal attention trained on 5 M video clips. The “future prediction” module enables the model to anticipate upcoming scene changes rather than react only to the current frame.

Data engineering : The pre‑training corpus totals 60 k hours of high‑quality robot experience, derived from 5 k hours of cleaned real‑robot data and 1 k hours of first‑person human manipulation video. Cleaning filters episodes with unreliable trajectories (using jerk, Z‑score of speed/acceleration) and misaligned video‑state pairs (checked via URDF‑based projection). A unified 55‑dimensional state‑action representation covers joints, end‑effector poses, gripper, head, waist and chassis signals, with missing dimensions padded. Automatic language annotation is performed by a VLM pipeline powered by Qwen3.6‑27B, generating atomic actions, interacted objects and concise commands for each sub‑task.

Benchmark performance : On the Shanghai Jiao Tong University GM‑100 suite, LingBot‑VLA 2.0 achieves higher average task‑progress scores and success rates than competitor models GR00T N1.7 and π0.5 on both the AgileX Cobot Magic and Galaxea R1 Pro dual‑arm platforms. In long‑range mobile tasks using the “Ark Arm + Songling” chassis and the StarDust S1 humanoid, LingBot‑VLA 2.0 also outperforms π0.5 in progress and success metrics, maintaining advantages across cross‑domain scenarios.

Efficient fine‑tuning and deployment : A post‑training version runs inference under 130 ms on an RTX 4090, dramatically lowering the cost and time required to adapt the model to new robot bodies, tasks, or environments. All model weights, training code, and the technical report (“From Foundation to Application: Improving VLA Models in Practice”) are released on Hugging Face, GitHub, ModelScope and the project website.

Conclusion : LingBot‑VLA 2.0 demonstrates that achieving robust cross‑embodiment VLA demands coordinated advances in sparse MoE architectures, spatial‑temporal supervision, and large‑scale heterogeneous data pipelines. The open‑source release provides a reproducible engineering stack that points to the next wave of embodied AI research.