AI-generated summary: When robots perform backflips en masse, it signals rapid progress in the cerebellum. But have you realized that the real bottleneck for getting robots to actually perform tasks has never been the cerebellum—it’s the brain? Ant Group’s newly open-sourced LingBot-VLA2.0 uses a single model to control 20 different robot configurations. For the first time, the industry is seriously accounting for the cost of repeated adaptation.Article author and source: AI New Era
Newzhongyuan reports
[New Intelligence Yuan Introduction] When robots performing backflips go viral, it signals rapid progress in the cerebellum. But have you realized that what has always bottlenecked robots from actually doing real work is not the cerebellum, but the brain? Ant Group’s LingBo has just open-sourced LingBot-VLA2.0, which uses a single model to control 20 different robot configurations. For the first time, the industry is seriously accounting for the cost of repeated adaptation.
Last year, the most talked-about scenes in the robotics world were all about the "body."
Robots can now perform backflips, robots can now complete a half-marathon…
Each video shows you: Look, how flexible its arms are, how steady its legs are, and how it can get back up after falling.
Lively, impressive, with thunderous applause.
However, this is merely a carefully orchestrated "Truman Show."
From a control theory perspective, this is typically a localized optimization achieved under specific closed environments, rather than genuine cognitive reasoning and model-driven long-term planning.
However, different configurations themselves represent a generalization challenge. Each robot brand and each hardware configuration is like an isolated linguistic island. We have built countless exquisite "bodies," yet we remain helpless before the "Tower of Babel" of how to make them speak and act.
The more impressive the demo, the more shocking the bill hidden behind it; the next client gets a new copy, and everything has to start over from scratch.
It is precisely within this industry gap—where passion runs hot but rationality is absent—that Robbyant, an embodied AI company under Ant Group, has chosen a radically different path:
They don’t chase trends by competing over hardware, specs, or physical strength, but instead target the industry’s most critical pain point—competing over whose intelligence is more general, practical, and capable of real-world application.
Since its inception, the LingBot-VLA series of large models has positioned itself as the "industry-wide general brain" in the field of embodied intelligence, aiming to solve the generalization challenge of "one brain, multiple machines" with a single unified brain.
Six months ago, Ant Lingbo open-sourced LingBot-VLA 1.0 using 20,000 hours of high-quality real-machine data, covering nine robot configurations.
Just now, Ant Group's Lingbo Technology open-sourced the embodied operation foundation model LingBot-VLA 2.0, directly addressing industry challenges.
Technical Report: From Foundation to Application: Enhancing VLA Models in Practice
Report address: https://github.com/robbyant/lingbot-vla-v2/blob/main/assets/LingBot_VLA_2_0.pdf
Project page: https://technology.robbyant.com/lingbot-vla-v2
Repository: https://github.com/robbyant/lingbot-vla-v2
Hugging Face: https://huggingface.co/collections/robbyant/lingbot-vla-v2
ModelScope: https://modelscope.cn/collections/Robbyant/LingBot-VLA-V2
Compared to the 1.0 era, LingBot-VLA 2.0 has undergone a comprehensive upgrade in its generalization capabilities.
Cross-platform dominance
From supporting one to supporting twenty
To cross the tipping point for large-scale deployment of embodied intelligence, models cannot merely succeed on a single robot or a single task; the real challenge is whether their capabilities can still be stably transferred after changing the embodiment, environment, or task.
In early 2026, Ant Group's Lingbo officially released and open-sourced the embodied operation foundation model LingBot-VLA 1.0.
At the time, the entire industry had serious concerns about the efficiency and cost of deploying VLA models in real-world settings—how could models run on actual dual-arm platforms with limited data and computing resources?
Version 1.0 provided the first-generation solution with an extremely pragmatic design.
LingBot-VLA 1.0 was the first to abandon the illusion of relying solely on simulated synthetic data, firmly committing to scaling up with large-scale physical data. During pre-training, it aggregated 20,000 hours of high-quality real-world manipulation trajectories from dual-arm robots. These data were not limited to a single configuration but spanned nine of the most popular dual-arm robot configurations at the time.
In terms of computational efficiency, the 1.0 team built an extremely efficient underlying training system. Under an 8-GPU FSDP/mixed-parallel training setup, their per-GPU pretraining throughput reached an impressive 261 samples per second, achieving 1.5x to 2.8x faster computation compared to mainstream VLA training frameworks at the time.
This high-throughput, lightweight architecture significantly reduces post-training costs. Users no longer need to deploy numerous GPUs for lengthy full-model retraining; instead, they can transfer bilateral manipulation capabilities to new task scenarios with just a few rounds of fine-tuning on downstream tasks.
In the 2.0 era, the pre-training phase covers over 20 robot configurations, spanning 17 major brands both domestically and internationally, including:
Leju KUAVO 4 Pro, Unitree, Astribot S1, AgileX, AgiBot G1/A2, Galaxea R1Pro/R1Lite, Galbot G1, Realman, Franka, ARX Lift2, Tian Gong, UR, GR2, MagicBot Gen1, Moz1, Zerith, Fuxi, Qinglong, and others.
The pretraining data has been expanded from 20,000 hours of dual-arm robot data to 50,000 hours of high-quality real-world robot data.
The list itself is highly impressive: rather than selecting just a few similar-looking robotic arms from a lab for a demo, it brings together a diverse range of mainstream humanoid, semi-humanoid, single-arm, and multi-composite robots with the greatest differences in physical configurations and kinematic constraints, achieving cross-embodiment generalization with a single set of large model weights.
The more critical change occurs at the scale of the action space.
LingBot-VLA 2.0 is the first to unify all body dimensions—including the head, waist, end effector (hand), and mobile base—into a single standardized action representation framework. The large model is no longer merely a “puppeteer controlling the arm,” but has learned to “move its entire body.”
This means it can not only skillfully manipulate objects on a tabletop with both arms, but also coordinate waist squats and base movement to perform large-scale 3D mobility, turning, and door opening, completing highly challenging long-horizon mobile manipulation tasks.
Based on the GM-100 dual-arm operation general benchmark, LingBot-VLA 2.0 demonstrates significant advantages across modalities in testing:
The GM-100 is a generalist benchmark specifically designed to challenge the generalization limits of robotic agents—one model, multiple tasks. In contrast, specialized benchmarks are designed for one model per task.
In terms of mobility, LingBot-VLA 2.0 has been preliminarily tested against π0.5 on two hardware platforms: the Ark Arm + Songling Chassis and the Stardust Intelligence Astribot S1.
The results show that LingBot-VLA 2.0 significantly outperforms others in task progress score and success rate on long-range manipulation tasks, with a consistent advantage in more challenging cross-domain scenarios, demonstrating superior capability in advancing long sequences of tasks and generalizing mobile manipulation.
In the mobile operation evaluation, the overall task is broken down into a series of consecutive sub-steps, each assigned a different score based on its difficulty and importance. The robot earns the corresponding score for each successfully completed step, and the final cumulative score reflects its actual progress capability in long sequential tasks.
Compared to evaluating solely based on final success rate, this step-by-step scoring approach provides a more detailed assessment of the model’s overall performance across diverse tasks such as movement, dual-arm collaboration, grasping, placing, opening doors, and cleaning.
From a technical perspective, this demonstrates the model’s cross-configuration generalization capability; from an industrial perspective, it means the repeated adaptation costs for different robot configurations can be reduced, enabling customers to validate new robots, tasks, and scenarios more quickly.
The Three Fundamental Technological Revolutions of LingBot-VLA 2.0
To enable the same brain to master twenty vastly different machine bodies, the technical approach can no longer rely on traditional "accumulation" strategies; it must depend on new advancements in model architecture and data pipelines.
Full-body coordination: Expanding from robotic arm control to full-body degree-of-freedom control
Many existing VLA large models have a limited action space.
Many models were originally designed with only a single robotic arm featuring six or seven degrees of freedom. Once the robot is equipped with a mobile base or transformed into a humanoid form with dexterous hands, the original action output layer becomes completely ineffective.
This results in the robot's body moving while its arms remain inert, or its arms grasping while its head and torso stand stiffly by—preventing smooth, coordinated full-body movement.
To overcome this physical limitation, LingBot-VLA 2.0 redesigns and standardizes the Unified Action Representation, encoding the movements of robots with various configurations into a 55-dimensional canonical vector.
Arm control: 14-dimensional joint positions for both arms + 14-dimensional end-effector (EEF) pose (supports 3D spatial position and quaternion-based rotation mapping).
End interaction: 2D gripper control or 12D dexterous hand joint control.
Torso and sensor control: 4D waist control + 2D head (view direction) control, enabling the head camera to real-time align with the operation center.
Mobile Control: 3D mobility base control.
Keep extension: 4D preserved space.
In real-world tasks, complex mobile operations are never mere combinations of actions.
It finally provides a system-level technical foundation for the elegant implementation of complex, full-body coordinated long sequences.
Generalized transition in three-dimensional spatial perception
Previous VLA models primarily relied on RGB camera images for end-to-end control. However, in real physical interactions, pure two-dimensional pixels easily lead to depth loss and parallax illusions.
When lighting changes slightly, the desk reflects glare, or a completely new cup appears, the robot cannot accurately determine whether it is 5 centimeters or 7 centimeters away from the gripper.
This innate defect at the retinal level directly causes the robot to frequently miss its grip in unstructured environments.
The hardcore technical solution in the paper: Dual-Query Distillation.
To inject precise spatial geometric intuition into large models without increasing real-time inference latency, LingBot-VLA 2.0 introduces LingBot-Depth for cross-modal feature distillation, enabling robots to better understand 3D space during operation.
The core value of this design lies in: during inference, the model requires no additional heavy depth estimation network and can autonomously "imagine" high-fidelity 3D spatial geometry from a single monocular RGB input within the latent space.
LingBot-VLA 2.0, trained on 60,000 hours of physically realistic data, achieves true cross-configuration transfer at its core. Whether the robot arm has seven or six degrees of freedom, or the base is wheeled or bipedal, the same set of motion commands can be seamlessly mapped onto entirely different hardware configurations after spatial perception alignment.
Introducing the "Future Prediction" task: The robot has finally gained a sixth sense.
The logic of traditional VLA is "reflex arc control."
The model observes the current frame and immediately outputs the joint torques for the current millisecond. This mechanism lacks the ability to predict future physical states.
The future prediction introduced by LingBot-VLA 2.0 centers on enabling the model to not only understand what it currently sees, but also predict what might happen next when generating actions.
Built on the DINOv3 backbone, Dino-Video in LingBot-VLA 2.0 introduces block-level causal temporal attention and 3D Rotational Position Encoding (3D-RoPE), training a native video representation model on 5 million video clips including general, first-person, and robotic footage.
On LARYBench, DINO-Video achieved the best performance on three out of four benchmarks, demonstrating its effectiveness as a robot-native temporal teacher model.
This means that LingBot-VLA 2.0 must first precisely predict the future motion of objects, the mechanical consequences of hand-object interactions, and the semantic evolution of the scene before generating an action.
Less mysticism, more engineering
The confidence behind LingBot-VLA 2.0 also comes from upgrades to its data processing pipeline and vast amounts of high-quality pretraining data.
They cleaned 50,000 hours of high-quality real-world data from approximately 100,000 hours of raw data, covering 20 different robot configurations and a wide range of tasks.
Single-arm: Franka, Flexiv;
Arms: AgileX, ARX Lift2, UR7e, Moz1;
Humanoid robots: AgiBot G1, Galaxea R1Pro/R1Lite, Astribot S1, Zerith H1, etc.
Humanoid robots: Unitree G1, Fourier GR-2, AgiBot A2, Leju KUAVO, etc.
Additionally, incorporate 10,000 hours of high-quality Ego data to enable the model to learn more natural interaction patterns and hand movement priors.
This isn't just about "stacking data." The new pipeline implements three rigorous filters:
- Smoothness and Stationarity Detection: Calculate the third-order difference (jerk) of motion/state and the Z-score of velocity/acceleration to remove segments with excessive jitter or prolonged stationary periods.
- Video - State Consistency Verification: Project the robot onto the image plane using URDF, perform manual and automated alignment checks, and filter out samples with blur, severe occlusion, frame drops, or multi-view inconsistencies.
- Ego data reconstruction: The VLM performs initial filtering, followed by SLAM to estimate camera trajectory and MANO for hand pose estimation, aligning hand movements into a unified world coordinate system. During training, the data is transformed back into the camera coordinate system using the current frame’s camera extrinsics to achieve a unified representation across data sources.
As a result, the same model can transfer its capabilities across 20 different ontologies, significantly reducing the cost of repeatedly adapting to different platforms. Previously, switching to a new robot was like “starting a new course”; now, it’s more like “putting on a new shell and continuing to use it.”
These designs significantly reduce post-training costs—building on v1’s achieved throughput improvement of 1.5–2.8x (261 samples/s per GPU), v2 further reduces the number of samples and computational resources needed to reach target performance through higher-quality data and a more optimized architecture. LingBot-VLA 2.0 is also open-sourcing a more efficient post-training version, with inference latency controlled to under 130 milliseconds on an RTX 4090.
Enterprise clients truly understand the math: it’s not about having larger model parameters, but about spending less money, reducing time, and supporting more robots.
The real industry pain point
Make robot capabilities reusable
When viewed within the industrial context, LingBot-VLA 2.0’s most significant value is not merely “another VLA model,” but rather its attempt to address a common industry-wide question:
Why has embodied intelligence not yet achieved scalability like large language models?
The answer is simple but harsh: language models operate through a uniform interface—text in, text out—while robots confront a fragmented world, where hardware, movements, sensors, tasks, and environments all differ.
Therefore, scaling embodied intelligence cannot rely solely on isolated capabilities; it requires three elements to be simultaneously in place.
First, the model can be reused across different blockchains; otherwise, onboarding each new client would be like starting from scratch every time.
Second, the action space must be uniformly expressed. Otherwise, the model learns only "muscle memory" specific to a particular machine, rather than general operational capability.
Third, the model must be able to predict how the world will change after an action. Otherwise, it can only react passively to the current frame and is likely to deviate during execution when faced with occlusions, moving objects, or long-sequence tasks.
LingBot-VLA 2.0 precisely addresses these three pain points.
Its goal is not to show off, but to reduce costs—not to make robots look more like magicians in demo videos, but to minimize the need for "starting over" at customer sites.
This is the true meaning of "the roll lands."
Benchmarking is certainly important—without metrics, there can be no comparison. But if a model can only win in a carefully controlled lab environment, it’s winning a leaderboard, not the real world of business.
For true embodied intelligence to enter factories, hospitals, nursing homes, households, and logistics, it must answer the three questions most important to employers: Do we need to retrain when replacing robots? Do we need to recollect data when changing scenarios? Can we reduce the failure rate of long tasks?
It is understood that Ant Lingbo, in collaboration with ecosystem partners such as Lujus and ecosystem clients like Guoda Pharmacy, has launched comprehensive commercial pilot tests in application scenarios including retail sorting, logistics sorting, and industrial use.
On the other hand, Ant Lingbo, in collaboration with data alliance partners such as Jianzhi Technology, is jointly building a standardized data system. An embodied intelligence ecosystem, centered on a cross-configuration VLA foundational model with deep participation from original equipment manufacturers and data institutions, is taking shape.
The answer from LingBot-VLA 2.0 is not the final outcome, but the direction is very clear: transition the bot's intelligence from "project-based manual customization" to "platform-based reuse."
Instantly follow ASI
⭐ Like, share, and click "see more" with one click ⭐
Star us to lock in instant updates from New智Yuan!
