A New Open-Source Robot AI Model Arrives — and It Speaks Every Robot's Language
Ant Group's robotics division, Robbyant, has released LingBot-VLA 2.0, an open-source robot AI model designed to control a wide range of robotic hardware using a single, unified artificial intelligence system. Released under the permissive Apache 2.0 licence, the model represents a significant step forward in so-called "cross-embodiment" robotics — the ability of one AI model to operate across radically different types of robot bodies without needing to be retrained from scratch for each one. For developers, IT architects, and policy professionals tracking the trajectory of open AI infrastructure, this release deserves close attention.
The model, which carries 6 billion parameters, was pretrained on roughly 60,000 hours of data — a dataset that combines 50,000 hours of robot trajectory recordings across 20 different robot configurations with 10,000 hours of egocentric human video. That scale places it among the largest openly available robotics foundation models to date. The decision to open-source the checkpoint under Apache 2.0 — one of the most developer-friendly licences available — means organisations can deploy, modify, and redistribute it without the licensing friction that often accompanies enterprise AI tools.
How Does One Model Control 20 Different Types of Robots?
The central technical challenge in cross-embodiment robotics is that different robot bodies have different "vocabularies" of movement. A robotic arm has entirely different degrees of freedom compared to a mobile platform with a humanoid torso. Historically, this has meant training a separate model for each hardware configuration — an expensive and time-consuming process that has slowed the deployment of AI-driven robots in real-world environments.
LingBot-VLA 2.0 solves this by mapping every supported robot configuration into a single 55-dimensional canonical action space. This unified representation covers robot arms, dexterous hands, waists, heads, and mobile bases. Essentially, the model learns to "think" about movement in a hardware-agnostic way, then translates that internal representation into the specific control signals each robot platform needs. For IT decision-makers evaluating robotics deployment at scale, this architectural choice has significant practical implications: one model to maintain, one model to audit, and one model to update across an entire fleet of heterogeneous hardware.
The approach draws on broader trends in foundation model design. Just as large language models like GPT and open-source alternatives such as Meta's LLaMA family learn general language representations that transfer across tasks, Robbyant's team has applied a similar philosophy to physical action. Research from institutions like Stanford's Human-Centered AI group and papers published on arXiv exploring generalised robot learning have long argued that cross-embodiment generalisation is the key bottleneck preventing AI robots from leaving controlled laboratory settings and entering messy, real-world environments.
The Architecture Under the Hood: Mixture-of-Experts Without the Overhead
One of the more technically interesting design decisions in LingBot-VLA 2.0 is its use of a token-level Mixture-of-Experts (MoE) action expert that operates without an auxiliary load-balancing loss. MoE architectures have become increasingly popular in large-scale AI models — they allow a model to scale its effective capacity by routing different inputs through different specialised "expert" subnetworks, rather than activating the entire network for every inference. The result is that the model can handle a wider variety of tasks and environments without a proportional increase in computational cost during inference.
What makes Robbyant's implementation notable is the removal of the auxiliary load-balancing loss typically required to prevent MoE models from routing everything through just one or two dominant experts. This simplifies training, reduces hyperparameter tuning complexity, and lowers the barrier for developers who want to fine-tune the model on their own hardware configurations. For open-source AI tools to be genuinely useful in enterprise and research contexts, training stability and reproducibility matter as much as raw benchmark performance.
The model also employs a technique called dual-query distillation, drawing supervision signals from two additional models: LingBot-Depth and DINO-Video. LingBot-Depth adds geometric understanding — essentially teaching the model to reason about the three-dimensional structure of the environment — while DINO-Video contributes temporal understanding, helping the model anticipate future states rather than simply reacting to the current frame. This "future-aware control" capability is particularly relevant for manipulation tasks where the robot needs to plan several steps ahead, such as assembling objects or handing items to a human collaborator.
"The shift toward open, cross-embodiment foundation models is arguably the most important structural change happening in robotics right now. A single model that generalises across hardware democratises deployment in a way that proprietary, hardware-specific systems simply cannot."
— Robotics AI researcher, commenting on the broader trend of generalised robot learning modelsBenchmark Results: How LingBot-VLA 2.0 Compares to Rivals
On the GM-100 generalist robotics benchmark — a standardised evaluation suite designed to test how well a model generalises across a wide range of manipulation tasks — LingBot-VLA 2.0 outperforms both π0.5, a competing generalised robot model, and its own predecessor, LingBot-VLA 1.0, across both evaluated platforms. While specific numerical scores were not disclosed in the initial release announcement, the performance margin on a widely recognised benchmark gives independent credibility to the model's capabilities beyond internal testing.
Benchmarks in robotics are notoriously difficult to standardise, and the field has historically struggled with reproducibility — a problem well documented in IEEE Spectrum's coverage of robot learning evaluation methodology. The use of GM-100 as a reference point is therefore significant, as it provides a common frame of reference for comparing models across different research groups. For IT professionals and enterprise buyers evaluating AI tools for physical automation, benchmark transparency is a prerequisite for trust — something the open-source release further supports by allowing independent replication.
| Model | Parameters | Training Data | Licence | GM-100 Performance |
|---|---|---|---|---|
| LingBot-VLA 2.0 | 6B | ~60,000 hours | Apache 2.0 | Best on both evaluated platforms |
| π0.5 | Not disclosed | Not disclosed | Proprietary | Outperformed by LingBot-VLA 2.0 |
| LingBot-VLA 1.0 | Not disclosed | Not disclosed | Apache 2.0 | Outperformed by LingBot-VLA 2.0 |
Why Open-Source Licensing Matters for AI Sovereignty and Enterprise Trust
The decision to release LingBot-VLA 2.0 under Apache 2.0 is not a minor footnote — it is arguably one of the most consequential aspects of the release for organisations thinking seriously about digital sovereignty and long-term AI strategy. Apache 2.0 permits commercial use, modification, and redistribution without requiring derivative works to remain open source. This makes it one of the most permissive licences available for AI models and places LingBot-VLA 2.0 in a fundamentally different category from proprietary robotics AI platforms offered as cloud services.
For European organisations in particular, where the European Commission's Open Source Software Strategy explicitly promotes open-source adoption as a mechanism for technological sovereignty, models like LingBot-VLA 2.0 offer a concrete path away from vendor lock-in. The ability to run inference and fine-tuning on-premises — rather than sending proprietary operational data to a third-party cloud API — directly addresses data sovereignty concerns that have become increasingly prominent under GDPR enforcement.
This is particularly relevant for manufacturers, logistics operators, and research institutions in regulated industries who are evaluating AI-driven automation but face strict data governance requirements. A model that can be deployed entirely within an organisation's own infrastructure, audited at the code level, and customised for specific hardware without engaging a vendor relationship, represents a materially different risk profile from a proprietary alternative.
What This Means for Developers Building on Open Robot AI Infrastructure
For developers working on robotics applications, the release of a well-documented 6B open-source robot AI model with a clear canonical action space is a significant infrastructure event. Rather than building proprietary motion planners for each hardware target, developers can now work with a single pretrained foundation and adapt it through fine-tuning — a workflow that will be familiar to anyone who has worked with large language models or vision-language models in software development contexts.
The 55-dimensional canonical action space also creates an interesting standardisation opportunity. If LingBot-VLA 2.0's action representation gains traction in the open-source community, it could become a de facto interface standard for robotic manipulation tasks — similar to how OpenAI's API schema influenced how developers structure LLM integrations. This kind of emergent standardisation is particularly valuable in hardware-adjacent software development, where the absence of common interfaces has historically fragmented developer ecosystems.
The dual-query distillation approach — borrowing geometric
Originally reported by MarkTechPost. Summarised and curated by European Purpose.
