# Robbyant Builds Out an Embodied-Native Full-Stack for Physical AI > Source: > Published: 2026-07-10 11:36:21+00:00 *In the span of a single week, Robbyant has released a cluster of models that, taken together, amount to something larger than any one of them on its own: a complete, ground-up stack for embodied artificial intelligence.* Robbyant, an embodied AI company within Ant Group, has spent the past week rolling out an unusually coordinated run of releases - LingBot-VA 2.0, LingBot-Depth 2.0 and LingBot-Vision, LingBot-VLA 2.0, and LingBot-World 2.0 alongside LingBot-Video. Announced one at a time, they read as incremental updates. Lined up together, they describe a deliberate strategy: six models spanning perception, reasoning, action, and world simulation, each built specifically for robots operating in the physical world rather than adapted from tools designed for digital content. That distinction is the thesis Robbyant keeps returning to, and it is worth taking seriously. Much of the current wave of embodied AI has been assembled by fine-tuning models originally built for video generation or language, then bending them toward robot control. Robbyant's argument is that this borrowing carries a hidden tax - models tuned for visual polish or fluent text were never optimized for causal prediction, physical accuracy, or real-time execution, and forcing the adaptation tends to erode generalization. Its answer is what the company calls an **Embodied-Native Full-Stack**: a set of components designed from scratch for physical intelligence, and increasingly released as open source. ## Teaching Robots to See Accurately The stack starts with perception. LingBot-Depth 2.0, a spatial perception model trained on 150 million samples, now tops 12 of 16 depth-completion benchmarks and roughly halves the depth error of its predecessor in the hardest indoor scenes, cutting RMSE from 0.132 to 0.062. Its most practical gains show up exactly where conventional depth cameras fail - glass, mirrors, and transparent surfaces, the everyday objects that quietly defeat most robots. Underneath it sits LingBot-Vision, a visual foundation model that the company says is the first to use "boundary structure" as a pre-training objective, reaching sub-pixel localization while training on only 160 million images, an order of magnitude fewer than comparable models. The efficiency claim matters as much as the accuracy one; it points to a stack designed to scale without runaway data costs. ## A Universal Brain, Trained on Real Robots Perception feeds action, and action is where LingBot-VLA 2.0 comes in. This vision-language-action model, now upgraded and open-sourced, is pitched as a "universal brain" - the piece the industry has struggled to standardize. Robbyant trained it on 60,000 hours of real-world physical data, drawn from 50,000 hours of cleaned real-robot interaction and 10,000 hours of first-person human manipulation. Just as important, that data spans 20 distinct robot body types from 17 manufacturers, including names like Unitree, AgiBot, Franka, and Fourier, covering single-arm, dual-arm, bipedal, and wheeled machines. The payoff is breadth: a single model meant to generalize across morphologies and coordinate whole-body movement across head, waist, hands, and a mobile base, rather than being retrained for every new chassis. ## World Models Built for the Physical World The most forward-looking layer is the world model, and here Robbyant makes its boldest claims. LingBot-VA 2.0, which the company describes as the industry's first embodied-native video-action world model, is built to predict how an action will change an environment and then choose the next step from that prediction; the demonstration of choice is a robot playing a real-time tabletop air hockey match against a person. LingBot-World 2.0, released open source under the name Infinity, pushes generation itself: hour-long continuous output at 720p and 60fps, held stable by a Causal Pretraining Paradigm and a proprietary attention mechanism the company calls MoBA, which it says eliminates the drift that usually degrades long-horizon generation. In its own hour-long stress tests, Robbyant reports zero quality loss. A native agent mechanism lets those generated worlds move from merely watchable to what it calls sustainably interactive, and a distilled fast-inference version, paired with LingBot-Video, streams frames as they are produced for a gaming-like experience on the company's Reactor platform. ## Why the Full-Stack Framing Matters Any one of these releases would be a respectable research update on its own. The more telling story is the assembly. By building perception, a control brain, and world simulation as a single embodied-native lineup, and open-sourcing much of it, Robbyant is doing something closer to platform-building than model-shipping. It is a recognizable move from Ant Group's wider playbook: release the components, invite developers in, and let an ecosystem settle around a common foundation. Whether the industry adopts Embodied-Native Full-Stack as a category or simply borrows the pieces, the wager underneath is clear. Robbyant is betting that the robots of the next few years will not be powered by digital AI wearing a physical costume, but by systems designed, from the first layer up, for the messy demands of the real world.