Ant LingBot-Depth 2.0: Scaling Robot Vision Training Data to 150 Million

On July 7, LingBo Technology, the embodied intelligence company under Ant Group, upgraded LingBot-Depth to version 2.0. This update revolves around a single core variable: the training data jumped from 3 million in the previous generation to 150 million, a full 50× increase. At the same time, the team also separated out the visual foundation model behind 2.0, LingBot-Vision, and linked it with Orbbec’s depth cameras for industry-side integration.

If last year’s version of LingBot-Depth looked more like a technical validation exercise (the first open-source model capable of handling transparent/reflective material depth perception without changing hardware), then 2.0 is clearly aimed at real-world deployment.

The Bottom Line First: This Is Not Just a Simple Version Increment

Anyone who has worked on robot vision knows that depth estimation has several longstanding pain points:

• Blurry edges: once object boundaries become fuzzy, grasp planning starts to wobble
• Small object loss: a charging cable or screw on a tabletop can disappear entirely in the depth map
• Long-range collapse: beyond 2 meters, errors grow nonlinearly
• Material sensitivity: transparent glass, reflective stainless steel, and light-absorbing black fabrics are the three major killers

The upgrade directions officially highlighted for LingBot-Depth 2.0 map exactly to these four issues: edge clarity, small-object recognition, long-range depth estimation, and robustness in complex scenes. They may look like four separate points, but fundamentally they all boil down to one thing — using larger-scale and more diverse data to push the model’s out-of-distribution generalization capability higher.

What Does 150 Million Samples Actually Mean?

A quick comparison makes the scale difference obvious. In academia, the commonly used NYU Depth v2 dataset has around 400,000 frames, while KITTI depth benchmarks are only in the tens to hundreds of thousands. Even large-scale open-source approaches like Depth Anything V2, which already pursued scale aggressively, only reach the tens of millions when combining annotations and pseudo-labels.

150 million is the result of throwing in every accessible data source: RGB-D cameras, structured light, ToF, stereo vision, and monocular pseudo-labels. LingBo has access to Ant’s application scenarios (financial service outlets, autonomous delivery, campus service robots), as well as sensor data accumulated by Orbbec, one of the leading depth camera manufacturers. At present, very few embodied AI companies in China can match this scale.

But data scale alone is not everything. Whether the marginal gains from moving from 3 million to 150 million can justify the training cost is another question. From the official comparison results, however, version 2.0 shows visibly improved performance over 1.0 in at least four hard scenarios — especially with small objects. Previously, a charging cable might disappear entirely in the depth map; now the model can render a continuous surface for it.

LingBot-Vision: From “Understanding” to “Precision”

Released alongside Depth 2.0, LingBot-Vision serves as the visual foundation of the system. LingBo defines the pipeline as moving “from understanding to precision”:

• Understanding: semantic comprehension, object recognition, scene classification — handled by the Vision layer
• Precision: geometric accuracy, distance, size, pose estimation — handled by the Depth layer

Neither side is new on its own, but the key point is that the depth model and visual foundation model share the same backbone. This means upper-layer tasks (such as grasping, navigation, and obstacle avoidance) no longer need to maintain separate feature pipelines. For robotics scenarios constrained by compute and sensitive to latency, this architectural integration has substantial value.

Compared with overseas approaches:

| Solution | Focus | Characteristics | |---|---|---| | Depth Anything V2 | Monocular depth | Strong generalization, weak on transparent/reflective materials | | Meta DINOv2 | Visual self-supervision | Strong semantics, geometry not specifically optimized | | Google RT-2/RT-X | End-to-end VLA | Directly outputs actions, depth handled implicitly | | LingBot-Depth 2.0 + Vision | Decoupled depth + vision | Modular and deployment-friendly for industry |

The route Ant has chosen essentially comes down to this: instead of building an end-to-end black box, it aims to thoroughly optimize and modularize the perception layer. For robots operating in deterministic scenarios like factories, warehouses, and services, this approach is easier to debug and deploy than end-to-end systems.

The Collaboration with Orbbec Is Worth Watching

Orbbec is one of China’s leading players in 3D vision sensors, with modules widely used in robot vacuums and humanoid robots. This collaboration with LingBo is fundamentally an “algorithm + hardware” integration — the sensor generates images, the model produces depth, and together they form a package that can be directly integrated into a robot controller.

This bundled approach is highly attractive for downstream robotics manufacturers. Previously, many humanoid and wheeled robot teams built their own perception stacks, and simply cleaning raw point clouds from RealSense or Orbbec sensors into usable depth maps could consume months of work from an entire perception algorithm team. With an out-of-the-box solution now available, at least the demo stage becomes much easier.

A Few Observations

In the current embodied AI wave, reading model papers alone is not enough. The factors that truly determine deployment velocity are three things: data, sensors, and deployment pipelines.

With this LingBot-Depth 2.0 upgrade, Ant has addressed the data side through sheer scale, partnered with Orbbec on sensors, and adopted an open-source + modular deployment strategy that is friendlier than end-to-end systems. From an industry perspective, this may be one of the most commercially valuable perception-layer updates in China’s embodied AI sector this year.

That said, some skepticism is still warranted:

• Few details have been disclosed about the distribution quality of the 150 million samples, and indoor scene overfitting remains possible
• No alignment benchmarks have yet been shown between the LingBot-Vision foundation model and mainstream VLMs (such as Qwen-VL or InternVL)
• End-to-end VLA approaches are advancing rapidly overseas, and it remains unclear how defensible the long-term moat of the perception-planning layered approach will be

For developers working on robotic arm grasping, AGV obstacle avoidance, or upper-layer planning for humanoid robots, it’s worth downloading and testing LingBot-Depth 2.0 — especially in environments where depth cameras traditionally struggle: glass doors, mirrors, black plastic parts. These scenarios previously required brute-force engineering tricks; now the model may be able to solve them directly.

One More Thing

Specialized embodied AI models like LingBot have not yet entered the API aggregation layer. But if your project needs both perception models and upper-layer planning or dialogue models such as GPT, Claude, Gemini, or DeepSeek, OpenAI Hub’s single-key access to all major models with OpenAI-format compatibility can save you the trouble of switching between multiple vendor SDKs. Direct domestic connectivity also makes it practical for teams building hybrid architectures that combine on-device robotics deployment with cloud inference.

References

• Baidu AI Studio - Technical Interpretation of the LingBot-Depth Model Technical details on the first-generation open-source LingBot-Depth model and its handling of transparent/reflective materials