Skip to main content

Ant Group's LingBot Trifecta: An Open-Source Embodied AI Stack from Video to World Models

In 2026, the race for embodied intelligence is intensifying, but the open-source community has consistently lagged behind—until Ant Group's Robbyant division delivered a stunning trifecta. From early 2026 to mid-year, LingBot-Depth, LingBot-VLA, and LingBot-World were open-sourced in rapid succession, forming a complete technical chain from spatial perception to decision-making to world simulation. The latest LingBot-World 2.0, codenamed Infinity, pushes world model generation into "hour-long" territory—an open-source first.

The Trifecta: Perception → Decision → Simulation​

Robbyant's LingBot series is not a collection of isolated breakthroughs but a carefully architected embodied AI technology stack.

LingBot-Depth serves as the "eyes" of the trifecta. This high-precision spatial perception model focuses on monocular depth estimation. In robot navigation and object grasping scenarios, accurate depth information is the foundation for all subsequent decision-making. Trained on large-scale diverse data, LingBot-Depth outputs high-quality depth maps across complex indoor and outdoor environments, providing robots with reliable three-dimensional spatial understanding.

LingBot-VLA is the "brain." VLA—Vision-Language-Action—represents the core paradigm of embodied intelligence today. It unifies visual perception, language understanding, and action planning into a single end-to-end model, enabling robots to comprehend natural language instructions and translate them into concrete physical action sequences. LingBot-VLA fills a critical gap in Chinese-language embodied foundation models, offering domestic robotics companies an autonomous and controllable model option.

LingBot-World is the "sandbox." This is a world model driven by video generation at its core, capable of producing physically consistent consecutive video frames based on input images and interaction signals. After its initial open-source release in January 2026, the team launched a major upgrade in July—LingBot-World 2.0, codenamed Infinity.

LingBot-World: From Video Generation to World Models​

LingBot-World takes a distinctive technical approach. Unlike traditional physics-engine-based simulators, it constructs world models through the lens of video generation. The intuition is straightforward: if a model can "imagine" what happens next, it understands the rules governing the world.

Built on the Wan2.2 architecture, the model supports both 480P and 720P resolutions. Its core technical highlights include:

High-fidelity diverse environments: LingBot-World maintains high fidelity and robust dynamics across realistic scenes, scientific contexts, cartoon styles, and beyond—not confined to a narrow training domain.

Minute-level long-term memory and consistency: The model sustains minute-level generation horizons while preserving contextual consistency across time—what the team calls "long-term memory." Where early video generation models exhibited drift and distortion after a few dozen frames, LingBot-World specifically optimizes against this.

Real-time interactivity: When generating at 16 FPS, system latency is under one second, meaning users can interact with the generated virtual world in real time via keyboard or control signals—moving forward, backward, or turning—and the world responds with immediate visual feedback.

Fully open-source: Model weights, code, and technical reports are all released under the Apache 2.0 license, hosted on HuggingFace and ModelScope. The team provides three model variants: Base (Cam) for camera pose control, Base (Act) for action signal control, and a Fast version for accelerated inference via KV caching.

Infinity: Entering the Age of Infinite Worlds​

In July 2026, Robbyant open-sourced LingBot-World 2.0, officially advancing the project to its Infinity stage. This is the only open-source world model to achieve hour-long, near-infinite generation duration across general domains.

Infinity's core breakthrough lies in solving the "drift accumulation" problem that plagues long-running world models. Traditional models, after generating hundreds of frames, gradually deviate from the initial scene, producing artifacts such as object disappearance and scene deformation. LingBot-World-Infinity overcomes this through an improved causal inference framework and chunked generation strategy, simultaneously achieving high dynamics, high visual quality, and semantic interaction during extended continuous generation.

This essentially provides embodied AI researchers with an "infinite sandbox": robots can engage in extended, uninterrupted training and testing without real-world physical constraints. For game developers and content creators, it means a real-time explorable, never-repeating AI-generated world.

The Open-Source Inflection Point for Embodied AI​

The strategic significance of the LingBot trifecta extends far beyond the three models themselves.

First, it demonstrates that major Chinese tech companies are transitioning from followers to leaders in embodied AI open-source. Previously, open-source embodied AI benchmarks were largely set by overseas teams. The completeness and engineering quality of the LingBot series signals a rapidly growing voice for Chinese teams in this arena.

Second, a full-stack "perception-decision-simulation" open-source framework is exceptionally rare. Most teams open-source only a single link in the chain, making end-to-end system-level research difficult for the community. Robbyant has connected the complete pipeline from visual input to action output to environmental simulation, dramatically lowering the barrier to entry for embodied AI research.

Finally, LingBot-World-Infinity's hour-long infinite world capability could represent a paradigm shift for embodied AI training. When simulation environments are sufficiently realistic and can run for arbitrary durations, robots can accumulate orders of magnitude more training data in virtual worlds than physically possible, accelerating sim-to-real knowledge transfer.

Looking Ahead​

The LingBot series remains in a rapid iteration phase. Current model capability boundaries—such as the accuracy of physical interactions and generalization of complex manipulations—still have substantial room for improvement. But the trifecta strategy has clearly revealed Robbyant's ambition: not to produce a single flashy demo, but to build foundational infrastructure that enables the entire embodied AI community to grow together. In this sense, the LingBot open-source stack may be the critical puzzle piece that moves embodied intelligence from the lab to real-world industry.