Ant Group / LingBot-VA
- inverse dynamics
- 2026
- Wan 2.2 (5B)
- World action model
Model cardPaper ↗
LingBot-VA is an autoregressive video-action model: it interleaves future-frame prediction and actions in one causal sequence on a Wan 2.2 backbone.
Key facts
Rows below are the register record in lib/models.ts. A dash means the value is not confirmed from a primary source. Classification follows Reuss, NVIDIA Technical Blog (15 June 2026) and the model's own paper.
| Name | LingBot-VA |
|---|---|
| Lab | Ant Group |
| Year | 2026 |
| Paradigm | inverse dynamics |
| Backbone | Wan 2.2 (5B) |
| Source | https://arxiv.org/abs/2601.21998 |
What the paper claims
- LingBot-VA is an autoregressive diffusion framework that learns frame prediction and policy execution together. Video and action tokens share a Mixture-of-Transformers, with a closed-loop rollout that can take ground-truth observations and an asynchronous pipeline that overlaps prediction with motor execution. Li et al., 2026
- The training corpus is approximately 16,000 hours of robot manipulation data across tasks and environments, including internally collected demonstrations. Li et al., 2026
- Reuss presents LingBot-VA as the modern inverse-dynamics counterpart to UniPi: a causal, closed-loop fine-tune of an open video backbone, rather than a from-scratch CNN video generator. Reuss, NVIDIA Technical Blog (15 June 2026)
What it evaluates
- The authors evaluate in simulation and the real world and report gains on long-horizon manipulation, post-training data efficiency, and generalization to novel configurations, including comparisons with π0.5. Li et al., 2026
On this register
Inverse dynamics. Generates the future video, then infers the actions that would produce it. That axis is defined in what is a world action model? and in Reuss, NVIDIA Technical Blog (15 June 2026).
Other inverse dynamics models on the register: UniPi, Video Prediction Policy, mimic-video, DVA.
The same backbone string (Wan 2.2 (5B)) also appears on Fast-WAM.