Inside LingBot’s Growing Open-Weight AI Ecosystem
A closer look at six models spanning vision, video, world generation, spatial reasoning, and embodied intelligence.LingBot AI models. AI generated imageMost AI models used in robotics were never originally designed for robots.Vision models are typically trained to recognize and classify images.…
A closer look at six models spanning vision, video, world generation, spatial reasoning, and embodied intelligence.LingBot AI models. AI generated imageMost AI models used in robotics were never originally designed for robots.Vision models are typically trained to recognize and classify images. Video models are optimized to generate visually convincing content. Language models are built to predict text. These systems can be adapted for robotics, but their internal representations may not capture the geometry, physical dynamics, and causal relationships needed to operate in the real world.Robbyant, an embodied AI company within Ant Group, is taking a different approach with its LingBot model family.Over the past several months, the company has introduced six models covering spatial perception, depth completion, video generation, interactive world simulation, and robot control:Lingbot VisionLingbot Depth 2.0Lingbot VideoLingbot World 2.0Lingbot VA 2.0Lingbot VLA 2.0These are not six versions of the same foundation model. Each one addresses a different layer of embodied intelligence. Together, however, they reveal a broader strategy for building AI systems that can perceive the physical world, predict how it will change, and act inside it.LingBot models. Image by Jim Clyde MongeThe idea behind embodied-native AIRobbyant describes the shared philosophy behind these models as “embodied-native.”Many robotics systems begin with a model pretrained for digital content. A video generator, for example, might learn from internet footage before being adapted to predict robot actions. This provides useful visual knowledge, but it also creates a mismatch between pretraining and deployment.A content-generation model is rewarded for producing realistic-looking video. A robot model must understand whether an action is physically possible, whether a task was completed, and how the environment will respond. Visual quality alone is not enough.An embodied-native model is built around these physical requirements earlier in its development. Depending on the model, this can influence the architecture, data mixture, training objective, temporal structure, and inference system.LingBot-Video includes robot-oriented footage covering manipulation, navigation, and egocentric perspectives. LingBot-VA 2.0 uses a causal architecture designed around action execution. LingBot-Vision focuses on dense spatial features rather than image-level recognition alone.This does not mean every LingBot model uses the same robotic dataset or training recipe. LingBot-Vision, for instance, is a self-supervised vision model. The common thread is that the models are designed with embodied applications in mind rather than treating robotics as a final adaptation step.The second idea is generalization.A robot policy that works with one arm in one laboratory is useful, but it is not a general foundation model. Robbyant is trying to build models that transfer across robot configurations, tasks, action spaces, and environments.LingBot-VLA 2.0 is the clearest example. Its pretraining data covers 20 robot configurations, including systems with movable heads, waists, mobile bases, and dexterous hands.Instead of learning one fixed hardware layout, the model is trained to handle multiple forms of embodiment.1. LingBot-Vision gives robots a spatial backboneMost vision models are trained to recognize what appears in an image. Robots also need to understand boundaries, surfaces, and the spatial relationships between nearby objects.LingBot-Vision is a family of self-supervised Vision Transformer backbones designed for dense spatial perception.LingBot models. Video from RobbyantAt a glanceUses masked boundary modeling to learn shapes, contours, and semantic regionsScales from ViT-S/16 to a 1.1B-parameter ViT-g/16Supports depth estimation, segmentation, object tracking, and depth completionDistills the Giant teacher into Large, Base, and Small deployment optionsBenchmark resultThe ViT-L model achieved an NYUv2 depth RMSE of 0.310, compared with 0.411 for DINOv2-L and 0.351 for DINOv3-L. Lower is better. It also remained within one percentage point of DINOv3-L on ImageNet linear classification.LingBot-Vision benchmark results. Image from RobbyantLingBot-Vision’s frozen features preserve object boundaries and geometric structure without losing high-level semantic information.Why should you care?A robot needs more than object labels. Boundary-aware features can improve grasping, navigation, segmentation, and other tasks that depend on precise object and surface localization.Learn more about this modelProject pageGitHub repositoryHugging Face collectionModelScope collectionTechnical paper2. LingBot-Depth 2.0 fills in missing geometryDepth sensors often fail around glass, mirrors, reflective floors, and distant surfaces. LingBot-Depth 2.0 reconstructs those missing measurements to produce more complete depth maps and point clouds.LingBot-Depth 2.0 examples. Image from RobbyantAt a glanceUses LingBot-Vision as its spatial-perception backboneComes in ViT-L/16 and ViT-g/16 configurationsExpands training data from 3 million to 150 million RGB-D samplesEvaluated across 14 benchmarks involving different cameras and missing-depth patternsBenchmark resultOn real sensor captures, LingBot-Depth 2.0 led six of eight configurations and recorded the strongest overall results on ClearGrasp, which focuses on transparent objects.LingBot-Depth 2.0 benchmarks. Image from RobbyantIt also achieved the best RMSE on seven of eight benchmarks involving block-masked and sparse input depth. The ViT-g version improved several results further, including:0.060 RMSE on block-masked DIODE-In2.298 RMSE on block-masked DIODE-Out0.124 RMSE on sparse Ibims-10.113 RMSE on sparse NYURaw depth sensors leave large gaps around difficult surfaces. LingBot-Depth 2.0 reconstructs them as more continuous geometry.Why should you care?Missing depth can cause a robot to misread glass or reflective surfaces as empty space. Better depth completion gives navigation and manipulation systems a more reliable geometric representation of the environment.Learn more about this modelLingBot-Vision paper with the Depth 2.0 technical sectionLingBot-Vision GitHub repositoryLingBot-Depth 1.0 project pageLingBot-Depth 1.0 GitHub repository3. LingBot-Video trains video models for physical intelligenceLingBot-Video is an open-source Mixture-of-Experts video foundation model designed specifically for embodied intelligence.It can generate videos from text and images, but that is only part of its purpose. The model can also simulate future trajectories, physical interactions, action progress, and task outcomes across different embodied systems.General and embodied video simulationLingBot-Video generates open-world scenes and future trajectories from text, images, and structured prompts. The model covers environments, materials, motion, and embodied tasks within the same video foundation model.Rather than producing only a visually plausible sequence, it can show how a scene develops as an agent moves, interacts with objects, or completes a physical task.LingBot-Video example output video. Video from RobbyantRobotics and embodied AIThe model covers robotic arms, humanoids, quadrupeds, mobile platforms, and egocentric agents. It can generate action progress, contact states, environmental feedback, and the final outcome of a task.For example, a robotics simulation can show an arm approaching an object, making contact with it, manipulating it, and responding to changes in the surrounding environment.LingBot-Video example embodied AI. Video from RobbyantMaterial and lighting propertiesLingBot-Video can simulate surface details, reflections, fluid motion, deformation, and changing lighting conditions. These properties remain visually consistent as the camera moves or objects interact.This is useful for physical simulation because contact does not affect every material in the same way. Fabric should fold, liquid should flow, and reflective surfaces should react consistently to movement and light.LingBot-Video example lighting capability. Video from RobbyantMotion and dynamicsThe model also covers human movement, animals, sports, egocentric footage, and natural dynamics. It aims to preserve coherent poses, realistic movement speeds, and stable motion across generated frames.These motion priors can help the model represent how bodies and objects are expected to move before it is adapted to a specific robotics task.LingBot-Video example motion and dynamics. Video from RobbyantAt a glanceUses a DiT-based Mixture-of-Experts architecture trained from scratchOffers a 1.3B dense model and a 30B-A3B MoE model with a refinerIncorporates more than 70,000 hours of embodied dataCovers manipulation, navigation, robots, and egocentric perspectivesReportedly provides around 3× faster inference through its MoE designBenchmark resultLingBot-Video scored 0.620 overall on RBench, ahead of Wan 2.6 at 0.607, Cosmos3 Super at 0.581, Seedance 1.5 Pro at 0.584, and Veo 3 at 0.563. It also recorded the highest listed scores for manipulation, long-horizon generation, and quadruped motion. SourceLingBot-Video benchmarking results. Video from RobbyantThe domain evaluation focuses on physical and embodied behavior rather than visual quality alone.Why should you care?Video models are increasingly being used as foundations for world models and robot policies. Training them to represent actions and physical outcomes could make those downstream systems more useful outside content generation.Learn more about this modelProject pageGitHub repositoryHugging Face collectionModelScope collectionTechnical paper4. LingBot-World 2.0 generates environments that keep evolvingInteractive world models must continue generating frames while responding to user or agent actions. Small errors can accumulate over time, causing textures, objects, and geometry to degrade.LingBot-World 2.0, also called LingBot-World-Infinity, is designed to generate explorable environments over an open-ended interaction horizon.What’s New in LingBot-World 2.0?If this is your first time hearing about LingBot-World, it is an interactive world model. You give it a starting image or a text prompt, and instead of producing a fixed video clip, it generates the world frame by frame as you move through it.Press W to walk forward. Swing a sword, and the model has to figure out what happens next. Think of it as a video game with no game engine underneath, just a neural network guessing the next frame based on your input.LingBot-World 2.0 demo. Image from RobbyantOne user can navigate as the player, while another acts as a director, guiding actions, events, or high-level intent.The first version could hold a scene together for a few minutes before quality drops. Version 2.0, which the paper calls LingBot-World-Infinity, makes a much bigger swing.Other key features of the modelUses causal pretraining to improve long-horizon consistencySupports real-time 720p generation at 60 frames per secondAdds actions such as attacking, archery, spell-casting, and shootingSupports text-driven environmental eventsUses pilot and director agents to control behavior and introduce new elementsHere’s an example video:https://medium.com/media/51352af688f3cf84e8cd12dd7c8ced08/hrefLingBot-World 2.0 generates persistent environments that respond to movement, actions, and text-driven events.Why should you care?World models could become controllable environments for games, agent evaluation, synthetic training data, and robot simulation. Long-horizon stability is one of the main barriers preventing those applications from moving beyond short demonstrations.Learn more about this modelProject page and demosGitHub repositoryHugging Face modelModelScope modelTechnical paperInteractive Reactor demoLearn more about LingBot Worlds 2.0 here:This Open-Source Model Lets You Explore AI Worlds in Real Time5. LingBot-VA 2.0 connects visual prediction with actionMost robot policies observe the current scene and immediately predict the next action. LingBot-VA 2.0 takes a different approach. It also predicts how the scene is likely to change after the robot acts.VA stands for video-action. The model learns future visual states and robot actions together, allowing it to connect an action with its expected physical outcome. It can imagine what may happen next, execute an action, and then update its prediction using the latest camera observation.LingBot-VA 2.0 was built from the ground up for robot control instead of adapting a video generator originally trained for content creation. Its semantic visual-action tokenizer places visual states and actions in a shared representation, while causal pretraining follows the same forward direction as real-world execution.LingBot-VA 2.0 demo. Image from RobbyantThis makes the model better suited to closed-loop control, where a robot must continuously compare its predictions with what actually happened.At a glanceUses a semantic visual-action tokenizer rather than a reconstruction-only VAEPretrains a causal DiT instead of converting a bidirectional video modelPlaces visual states and latent actions inside a shared semantic spaceUses sparse MoE layers to increase capacity without activating every parameterRe-grounds predictions using the latest real camera observationCan adapt from 10 to 15 demonstrations and supports some zero-shot transfersBenchmark resultLingBot-VA 2.0 achieved an average success rate of 93.6% on RoboTwin 2.0, including 93.8% in clean environments and 93.4% in randomized environments. The previous LingBot-VA averaged 92.2%, while Motus reached 87.9%. SourceIts acceleration pipeline also reduced inference from 927 ms to 142 ms per chunk, increasing the reported peak asynchronous control frequency from 35 Hz to 225 Hz.LingBot-VA 2.0 system architecture. Image from RobbyantLingBot-VA 2.0 predicts future visual states while the robot executes its current actions, then corrects the next rollout using a fresh observation.Why should you care?Robot policies must react to what actually happened, not only what the model expected to happen. Closed-loop re-grounding allows the system to update its plan when objects move, actions fail, or the environment changes.Learn more about this modelProject pageTechnical paper on arXivLingBot-VA GitHub repositoryNote: The public repository and available Hugging Face checkpoints still primarily document the first-generation LingBot-VA. The VA 2.0 technical report is public, but a separate VA 2.0 checkpoint was not listed during this review.6. LingBot-VLA 2.0 generalizes across robots and tasksLingBot-VLA 2.0 focuses on a related but different problem. VLA stands for vision-language-action. The model receives visual observations and language instructions, then translates them into robot actions.A typical VLA model may work well on the robot configuration and tasks included in its training data. Performance can drop when the hardware, camera position, action space, or environment changes. Robbyant designed LingBot-VLA 2.0 to improve this type of cross-embodiment generalization.The model was pretrained on 50,000 hours of robot trajectories spanning 20 configurations, along with 10,000 hours of egocentric human video. These configurations cover single-arm and dual-arm systems, humanoids, mobile platforms, dexterous hands, movable heads, and articulated waists.LingBot-VLA 2.0 unified action space. Image from RobbyantLingBot-VLA 2.0 maps these different bodies into a shared 55-dimensional action representation. This allows the model to learn common manipulation patterns while still supporting hardware-specific controls. It also uses depth and video representations to reason about current geometry and how the scene may change after an action.At a glanceUses approximately 60,000 hours of pretraining dataIncludes 50,000 hours of trajectories across 20 robot configurationsAdds 10,000 hours of egocentric human videoSupports arms, end-effectors, grippers, hands, waists, heads, and mobile basesMaps different embodiments into a shared 55-dimensional action representationUses depth and video distillation to learn current geometry and future dynamicsBenchmark resultOn the GM-100 generalist benchmark, LingBot-VLA 2.0 reached a 66.2% progress score and 34.4% success rate on AgileX Cobot Magic. Version 1.0 recorded 58.2% and 30.0%.On long-horizon mobile manipulation, it achieved:77.1 / 60.0 for in-domain refrigerator sorting37.0 / 13.3 for out-of-distribution refrigerator sorting84.3 / 66.7 for in-domain stove cleaning67.5 / 40.0 for out-of-distribution stove cleaningThe values represent progress score and success rate.LingBot-VLA 2.0 unified action space. Image from RobbyantLingBot-VLA 2.0 learns from 20 robot configurations rather than tying its foundation model to one hardware design.Why should you care?Collecting a large dataset for every robot is expensive. A model that transfers useful behavior across bodies, tasks, and environments could reduce the amount of robot-specific training required before deployment.Learn more about this modelProject pageGitHub repositoryHugging Face collectionModelScope collectionTechnical paperHow the six models fit togetherThe LingBot releases are better understood as complementary models than as a single end-to-end product.How LingBot’s six models fit together. Image by Jim Clyde MongeNot every model feeds directly into every other model. However, several technical connections are already visible.LingBot-Vision provides the encoder for LingBot-Depth 2.0. Video and depth representations provide both semantic and geometric cues for predictive training in LingBot-VLA 2.0. LingBot-Video provides data and dynamics priors used by LingBot-VA 2.0.Together, the releases cover perception, geometry, dynamics, simulation, and action. They also show how Robbyant is trying to develop those capabilities around a shared understanding of embodied intelligence.Final thoughtsIt is pretty interesting to see how well Chinese models have been doing over the past couple of months. Some are already close to the best US models, and many of them are open weights. Most of the big US models, meanwhile, are still closed.Props to AI labs like Robbyant for building models this powerful and making most of them available for people to download, test, and improve. LingBot-Depth 2.0 is not open yet, but the rest of the lineup already gives developers a lot to experiment with.I can see these models being useful in computer vision, AR, video simulation, and robotics. I am curious to see what people build when they start connecting the different LingBot models together.Which LingBot model interests you the most? Let me know in the comments.This story is published on Generative AI. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories.Subscribe to our newsletter and YouTube channel to stay updated with the latest news and updates on generative AI. Let’s shape the future of AI together!Inside LingBot’s Growing Open-Weight AI Ecosystem was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.Source: Generative AI Pub — Published — Category: Image AI