Ant Group unveils LingBot-VLA 2.0 for cross-embodiment robot control

Ant Group has released LingBot-VLA 2.0, an embodied model focused on turning VLA policies from single-robot control into a more general cross-embodiment policy. The update matters because discussion around it spans not just model scale, but also hardware coverage, inference practicality, and where current “generalist” robot policies still fall short.

Key details

According to the posts, LingBot-VLA 2.0 was trained on about 60,000 hours of data: 50,000 hours of real robot trajectories and 10,000 hours of first-person human video. Pretraining covers 20 robot configurations, ranging from single-arm and dual-arm systems to mobile bases and humanoids, with examples including Unitree G1 and Fourier GR-2.

Method-wise, the model predicts next-step depth and semantic features as a prior for task progress understanding. Its action space goes beyond arm-and-gripper control to also include the head, waist, and mobile base, aiming to let one policy interface adapt across different hardware forms.

Results and usability

Posts say that under a generalist setup across 9 GM-100 tabletop tasks, LingBot-VLA 2.0 performed better overall than π₀.₅ and GR00T N1.7 on two robot platforms. It was also described as leading on long-horizon mobile tasks.

On deployment details, Robbyant has open-sourced the post-training code. Inference is said to run on a single NVIDIA GeForce RTX 4090D at about 130 ms with 10 denoising steps, emphasizing that the model is being positioned not only as a research result but also as something closer to practical use.

Limits and open questions

At the same time, omarsar0 noted that LingBot-VLA 2.0 still shows uneven performance across tasks in general control. In that reading, the model is stronger in mobility-heavy and long-horizon scenarios, but there is still no single policy that can stably cover all manipulation tasks. That leaves the release as a notable step toward more realistic embodied deployment, while also underscoring the remaining engineering and capability gaps before one-policy-fits-all robotics becomes real.

2026-07-15 ~ 2026-07-16 · 8 related posts