Robbyant open-sources robot model for multiple types

Robbyant has open-sourced LingBot-VLA 2.0, a vision-language-action model for robots designed to work across a range of robot types without retraining.

The release expands the company's embodied AI software portfolio, following separate launches in its LingBot 2.0 series focused on robotic perception and vision.

Robbyant, an embodied AI business within Ant Group, said the latest model was trained on 60,000 hours of real-world physical data, including 50,000 hours of cleaned robot interaction data and 10,000 hours of distilled first-person human manipulation data.

The training set drew on 20 robot morphologies from 17 manufacturers, including Leju, AgiBot, Unitree, AgileX, Galaxea, Galbot, Astribot, RealMan, Franka, ARX, X-Humanoid, Fourier, MagicLab, Spirit AI, Zerith, Flexiv and Qinglong.

The model covers single-arm, dual-arm, bipedal and wheeled robot designs. It also extends control to the head, waist, hands and mobile chassis, allowing one system to manage more of a robot's body during tasks.

Benchmark results

Robbyant's performance claims focused on cross-morphology testing, which has become increasingly important as robotics developers try to avoid building separate software stacks for different machines. Many existing models remain tied to specific robots and tasks, raising deployment costs when customers switch hardware or expand into new environments.

On Shanghai Jiao Tong University's GM-100 benchmark for dual-arm manipulation, Robbyant said LingBot-VLA 2.0 recorded higher average task progress scores and success rates than π0.5 and GR00T N1.7 on AgileX Cobot Magic and Galaxea R1 Pro platforms.

In long-horizon mobile manipulation tests on the ARX Arm with AgileX Chassis and on the Astribot S1 platform, the model also exceeded π0.5 on both task progress scores and success rates, according to the company.

If replicated more broadly, those results would suggest that a single model can transfer more readily between machines with different structures and movement patterns. That remains a central issue for robotics developers trying to move from demonstrations to wider commercial use.

Deployment focus

Robbyant also highlighted deployment efficiency as a key part of the release. The company said LingBot-VLA 2.0 includes a version tuned for post-training efficiency, with latency below 130 milliseconds on an RTX 4090.

That matters because post-training and hardware adaptation are often among the most expensive stages of putting robotics software into production. Companies across the sector have been trying to reduce the amount of task-specific and machine-specific retraining required before systems can be used in warehouses, shops or factories.

Robbyant said it is already testing the model in commercial pilot projects in retail sorting, logistics and industrial automation. Partners in those trials include hardware groups Leju and Ti5 Robot, and enterprise customers GuoDa Drugstore and Longsheng Technology.

It is also working with GenRobot.ai to build standardised data ecosystems, reflecting a wider industry push to assemble larger and more consistent robotics datasets for training models across different hardware platforms.

Wider model push

The open-source release of LingBot-VLA 2.0 comes alongside other recent Robbyant model announcements in embodied AI. Earlier in the week, the company introduced LingBot-Depth 2.0, a spatial perception model, and LingBot-Vision, a visual foundation model.

Those models address a separate robotics challenge: how machines interpret physical surroundings with enough precision to operate safely around objects, surfaces and people. Robbyant said LingBot-Depth 2.0 was trained on 150 million samples and ranked first in 12 of 16 depth completion benchmarks, while LingBot-Vision was trained on 160 million images.

The company added that LingBot-Depth 2.0 reduced root mean square error in demanding indoor scenarios from 0.132 to 0.062 compared with its predecessor. It also performed strongly on transparent and reflective surfaces, known problem areas for conventional depth cameras, according to Robbyant.

For commercial perception applications, Robbyant said it has established a collaboration with Orbbec. The companies are working on products that combine Orbbec camera hardware with Robbyant's spatial perception software, including an SDK for robotics customers using Gemini 330 series cameras.

Taken together, the VLA, depth and vision releases show how Robbyant is trying to build a broader software stack for embodied AI rather than a single-purpose robotics model. The strategy places it in a fast-growing field where companies are racing to pair robot hardware with reusable software that can handle perception, planning and action across varied settings.

LingBot-VLA 2.0 is fully open-sourced and is undergoing pilot testing in retail sorting, logistics and industrial automation scenarios, Robbyant said.