# Is LingBot-VLA 2.0 the Most Broadly Trained Open VLA Model for Humanoids?

Robbyant, the embodied AI division of Ant Group, today open-sourced LingBot-VLA 2.0 — a [vision-language-action model](https://humanoidintel.ai/glossary/vision-language-action-model) pre-trained on 60,000 hours of physical interaction data spanning 20 distinct robot morphologies from 17 manufacturers. The model, which succeeds LingBot-VLA 1.0 released in January 2026, claims benchmark superiority over [Physical Intelligence (π)](https://humanoidintel.ai/companies/physical-intelligence)'s π0.5 and Nvidia's GR00T N1.7 on Shanghai Jiao Tong University's GM-100 dual-arm benchmark, while delivering a 3x improvement in inference efficiency with latency held under 150 milliseconds. Those are meaningful numbers if the evaluation methodology holds up to scrutiny — and that's the key question the broader industry will need to answer. The training corpus draws from 50,000 hours of cleaned real-robot interaction data and 10,000 hours of distilled first-person human manipulation data, covering single-arm, dual-arm, bipedal, and wheeled configurations. Commercial pilots are already underway with hardware partners [Leju Robotics](https://humanoidintel.ai/companies/leju-robotics) and Ti5Robot, and enterprise customers GuoDa Drugstore and Longsheng Technology in retail sorting, logistics, and industrial environments.

---

## What LingBot-VLA 2.0 Actually Does Differently

The core technical argument Robbyant is making is morphological generalization at scale. Most VLA models today are trained on a narrow slice of robot embodiments — typically a single arm configuration, occasionally dual-arm. LingBot-VLA 2.0 explicitly claims coverage across 20 morphologies sourced from manufacturers including [AGIBot](https://humanoidintel.ai/companies/agibot), [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics), [Astribot (Stardust Intelligence)](https://humanoidintel.ai/companies/astribot), [Fourier Intelligence](https://humanoidintel.ai/companies/fourier-intelligence), [MagicLab](https://humanoidintel.ai/companies/magiclab), [Galaxea AI](https://humanoidintel.ai/companies/galaxea-ai), [Galbot](https://humanoidintel.ai/companies/galbot), Leju, Franka, ARX, RealMan, X-Humanoid, Spirit AI, Zerith, Flexiv, AgileX, and Qinglong.

That breadth matters for a reason beyond benchmark optics. The long-standing complaint about deploying VLA models in commercial settings is that each new hardware platform requires expensive fine-tuning cycles. A model that genuinely generalizes across morphologies — including bipedal configurations relevant to humanoid platforms — would dramatically compress that post-training cost.

The DoF expansion is also notable. LingBot-VLA 2.0 now incorporates head, waist, [end-effector](https://humanoidintel.ai/glossary/end-effector) (hands), and mobile chassis control into a unified policy, enabling coordinated [whole-body control](https://humanoidintel.ai/glossary/whole-body-control). For humanoid deployments specifically, this is the architectural direction that matters: a policy that can only control arms is commercially limited when the platform is a full-body system expected to navigate and manipulate simultaneously.

---

## The Benchmark Claims Deserve Skepticism

The headline performance numbers — outperforming π0.5 and GR00T N1.7 on SJTU's GM-100 benchmark — are significant claims that warrant careful reading.

First, GM-100 is a specific dual-arm benchmark. Task progress scores and success rates on a defined benchmark set are not equivalent to general-purpose capability in unstructured environments. Robbyant's own source material does not disclose the specific numerical scores, only that LingBot-VLA 2.0 achieved "leading" results and "outperformed" competitors — which makes independent verification impossible from this announcement alone.

Second, the test platforms cited — AgileX Cobot Magic and Galaxea R1 Pro for dual-arm tasks, ARX Arm + AgileX Chassis and Astribot S1 for mobile manipulation — are all platforms already represented in the training data. The more demanding test of cross-morphology generalization would be zero-shot performance on hardware not seen during pre-training. That result isn't reported here.

Third, the human manipulation data pipeline — 10,000 hours of "distilled first-person human manipulation data" — raises questions about the distillation methodology. How that data is aligned to robot action spaces is critical to whether it actually transfers, and the announcement is silent on this.

None of this invalidates the work. But investors and engineering teams evaluating LingBot-VLA 2.0 for deployment should treat the benchmark claims as a starting point for evaluation, not a conclusion.

---

## The Deployment Play: Why the 150ms Latency Number Matters

The 3x inference efficiency improvement, with latency capped under 150 milliseconds, is arguably more commercially relevant than the benchmark results. Real-time control loops for manipulation tasks typically require sub-200ms inference to feel responsive and avoid compounding errors in long-horizon tasks. Hitting sub-150ms means LingBot-VLA 2.0 can plausibly run on-device or on edge compute without requiring cloud round-trips — a major barrier to commercial deployment that has quietly slowed adoption of earlier VLA architectures.

The commercial pilot structure is also worth noting. Robbyant is testing in retail sorting and logistics alongside GuoDa Drugstore and Longsheng Technology — environments with relatively constrained task variation but high throughput requirements. These are exactly the conditions where a fast, morphologically flexible model has the clearest near-term ROI argument. Conversely, they are not the unstructured household or industrial assembly environments where VLA models tend to fail in ways benchmarks don't capture.

---

## Industry Trajectory: Open-Source as Competitive Strategy

The open-source release is a deliberate ecosystem play. By open-sourcing LingBot-VLA 2.0, Robbyant — operating within Ant Group's financial and infrastructure umbrella — is positioned to attract hardware partners who need a capable foundation model but lack the data resources to train one from scratch. The partnership with GenRobot.ai to build standardized data ecosystems suggests the longer-term strategy: establish LingBot as the data and model standard, then monetize at the infrastructure and deployment layer.

This mirrors the approach Physical Intelligence has taken with selective open releases, and reflects a broader pattern in the VLA space where model weights are increasingly commoditized while data pipelines and deployment tooling become the defensible moat.

For the humanoid robotics ecosystem specifically, the significance is this: a well-validated, open, cross-morphology VLA model lowers the barrier for smaller hardware startups to field capable policies without building their own foundation model from scratch. If LingBot-VLA 2.0's benchmark claims hold up under independent evaluation, it could meaningfully accelerate time-to-deployment for bipedal platforms that currently lack the data infrastructure to train at this scale.

---

## Key Takeaways

- **60,000 hours of training data** — 50K hours real-robot, 10K hours human first-person — across 20 morphologies from 17 manufacturers, making it one of the broadest training corpuses publicly disclosed for a VLA model.
- **Claimed benchmark superiority** over π0.5 and GR00T N1.7 on SJTU's GM-100 dual-arm benchmark, though specific scores are not disclosed and all test platforms appear in the training data.
- **3x inference efficiency improvement** with sub-150ms latency, making real-time edge deployment more viable for commercial applications.
- **Whole-body control** now includes head, waist, hands, and mobile chassis — architecturally relevant for full humanoid deployments.
- **Open-source release** signals an ecosystem-building strategy, with commercial pilots already running in retail and logistics environments.
- Independent, out-of-distribution evaluation will be the real test of the cross-morphology generalization claims.

---

## Frequently Asked Questions

**What is LingBot-VLA 2.0 and who makes it?**
LingBot-VLA 2.0 is a vision-language-action model developed by Robbyant, an embodied AI company operating within Ant Group. It was released as an open-source upgrade to LingBot-VLA 1.0, which launched in January 2026.

**How does LingBot-VLA 2.0 compare to π0.5 and GR00T N1.7?**
According to Robbyant, LingBot-VLA 2.0 outperforms both π0.5 from Physical Intelligence and Nvidia's GR00T N1.7 on Shanghai Jiao Tong University's GM-100 benchmark across dual-arm and mobile manipulation tasks. However, specific scores are not disclosed in the announcement, and independent third-party validation has not yet been published.

**What robot platforms is LingBot-VLA 2.0 compatible with?**
The model was pre-trained on data from 20 robot morphologies spanning 17 manufacturers, including single-arm, dual-arm, bipedal, and wheeled configurations. Named platforms include those from AGIBot, Unitree, Astribot, Fourier, Galaxea, Galbot, Leju, Franka, AgileX, and others.

**What does the 150ms latency figure mean for deployment?**
Sub-150ms inference latency means the model can support real-time manipulation control loops without requiring cloud compute, enabling edge deployment on commercial hardware — a key requirement for industrial and retail environments where network dependency creates reliability risks.

**Is LingBot-VLA 2.0 suitable for humanoid robots specifically?**
The model's expanded DoF support — covering head, waist, hands, and mobile chassis — and its inclusion of bipedal robot morphologies in training data make it architecturally relevant for full humanoid platforms. However, most reported benchmarks focus on dual-arm and mobile manipulation tasks rather than bipedal locomotion-integrated scenarios.