Ant's LingBot-VA 2.0 Drives Robots by Predicting the Video of the Next Move, at 225 Hz

TL;DR

Robbyant, Ant Group's embodied-AI unit, announced LingBot-VA 2.0 today: a roughly 15.3B-parameter mixture-of-experts world model that controls robots by predicting what the camera will see next, and that the company claims is the first such model pretrained natively for acting rather than adapted from a video generator. The paper reports 93.6% average success on RoboTwin 2.0's 50 bimanual manipulation tasks, 13.8 points ahead of Physical Intelligence's pi-0.5, plus adaptation to new real-world tasks from 10 to 20 demonstrations and a peak 225 Hz control frequency from an asynchronous inference trick worth understanding. The paper is public and much of the surrounding LingBot stack is already on Hugging Face. The 2.0 weights themselves are not up yet.

Who shipped what

Robbyant has spent the past week emptying its lab onto the internet: an open 6B vision-language-action model (lingbot-vla-v2-6b) on July 8, a 14B causal world model (lingbot-world-v2-14b-causal-fast) on July 9, and now LingBot-VA 2.0 as the capstone. The paper, titled Native Video-Action Pretraining for Generalizable Robot Control, hit arXiv on July 9 and the announcement landed July 11.

The marketing line is "the industry's first embodied-native video-action world model." Industry-first claims in robotics have the shelf life of a banana, but the architectural argument underneath is concrete and checkable.

Why "native" is the whole argument

Most video-action models until now started life as a text-to-video or image-to-video generator, then got surgery: bolt on an action head, fine-tune on robot data, hope the video prior transfers. That inheritance brings two problems. The generator's attention is bidirectional (it was trained to see the whole clip at once, which a robot acting in real time cannot), and its tokenizer is a reconstruction VAE optimized for pretty pixels rather than for the physics that matter to a gripper.

LingBot-VA 2.0 replaces both. A semantic visual-action tokenizer (a ViT autoencoder with 96 latent channels, aligned to frozen vision foundation models) encodes what the scene means rather than how it shimmers, and the backbone is a causal diffusion transformer pretrained from scratch to predict forward in time, with a sparse mixture-of-experts video trunk. It is the difference between hiring a film-school graduate and retraining them to drive a forklift, versus training the forklift driver from day one and letting them pick up some cinematography on the job.

The numbers, and the one that matters

On RoboTwin 2.0, the paper reports 93.8% success in the clean setting and 93.4% under domain randomization, for a 93.6% average. Baselines in the same evaluation: LingBot-VA 1.0 at 92.2%, Motus at 87.9%, pi-0.5 at 79.8%, and X-VLA at 72.9%.

The number that matters is the gap between clean and randomized: 0.4 points. Policies that ace a tidy benchmark and collapse the moment the lighting changes are the oldest failure mode in robot learning, so a model that barely notices the randomization is the more interesting claim than the headline score.

For the real world, Robbyant ran four tasks (fruit sorting, pen collection, drawer tidying, plate handover) with 20 teleoperated demonstrations per task, and reports beating both pi-0.5 and VA 1.0 on success rate and task progress. The paper also claims adaptation to new tasks from as few as 10 to 15 demos, which, if it holds up outside Ant's lab, is the practical headline for anyone who has ever paid for teleoperation data.

The 225 Hz trick: acting on a future that hasn't happened yet

World models have an obvious problem as controllers: generating video is slow, and a robot arm cannot wait politely mid-reach. LingBot-VA 2.0's answer is an asynchronous scheme the paper calls foresight reasoning. While the arm executes the current action chunk, the model is already predicting the latent state of the world at the end of that chunk and planning the next one against its own prediction, then re-grounding on the actual camera frame each cycle so errors do not compound.

It is how you drive a car: you steer toward where the road will be when you get there, not where it was in the last frame you processed, and you glance up constantly to correct. That overlap is what lifts peak closed-loop execution to 225 Hz despite the model's size.

Inside the 15.3B: most of it is asleep

The parameter budget breaks down as a 13.0B video backbone, roughly 1.7B of multi-chunk prediction heads, and a 0.6B action expert. Thanks to the sparse MoE routing, only about 2.5B parameters are active per token at inference, which is how a world model this large gets within shouting distance of real-time control at all.

What you can actually download today

The honest availability picture, as of this writing:

• LingBot-VA 1.0: code and weights are Apache 2.0 on GitHub and Hugging Face (base plus RoboTwin and LIBERO-LONG post-trains). The repo cites roughly 24GB of VRAM for single-GPU RoboTwin evaluation in offload mode, so a 3090 or 4090 homelab box is in play.
• The v2 stack around it: the 6B VLA, the 14B causal world model, and video backbones up to a 30B-A3B MoE went up on Hugging Face this week.
• LingBot-VA 2.0 itself: paper and project page only. The weights are, in the proud frontier-lab tradition, coming.

Caveats, straight-faced

Every number above is vendor-reported; no third party has replicated the RoboTwin results yet. The real-world comparisons against pi-0.5 are published as charts rather than tables, so exact per-task success rates are not independently quotable. And "first embodied-native" is Robbyant's framing of its own design choice, not an adjudicated fact.

The bigger open question is the release itself. Until the 2.0 weights and license land, this is a strong paper attached to a very open 1.0, and the gap between those two things is where robotics announcements have historically gone soft.

Key Takeaways

• Ant Group's Robbyant announced LingBot-VA 2.0, a ~15.3B MoE video-action world model it claims is the first pretrained natively for robot control rather than adapted from a video generator.
• Vendor-reported results: 93.6% average on RoboTwin 2.0's 50 bimanual tasks, ahead of pi-0.5 (79.8%) and its own 1.0 (92.2%), with only a 0.4-point drop under domain randomization.
• An asynchronous "foresight reasoning" scheme plans the next action chunk against the model's own prediction of the near future, reaching a peak 225 Hz in closed-loop control.
• New real-world tasks were learned from 10 to 20 demonstrations each, the practical claim that matters most if it replicates.
• VA 1.0 is fully open (Apache 2.0, ~24GB VRAM to evaluate) and the surrounding v2 stack is on Hugging Face, but the 2.0 weights are not public yet.

Sources: arXiv 2607.08639, robbyant/lingbot-va (GitHub), Robbyant on Hugging Face, MarkTechPost