A new open-source AI model for robots is putting 3D understanding at the center of the push toward more capable automation. Developed by researchers at the Institute for Computer Science, Artificial Intelligence and Technology (INSAIT) in Bulgaria, SPEAR-1 is designed to act as a brain for industrial robots as they grasp, move, and manipulate objects.
The release matters because robot intelligence remains difficult to generalize. A robot arm can be trained to perform a narrow task, but changes to the arm, the object, or the surrounding environment can force retraining from scratch. SPEAR-1 aims at that gap by giving researchers and startups a more open way to experiment with models that better understand the physical world.
Why 3D data changes the problem
Many robot foundation models are built on top of vision language models, or VLMs. These systems can connect images and language in useful ways, but their view of the world is limited by training that often comes from labeled 2D images.
Robots do not work inside flat pictures. They operate in physical space, where objects have position, depth, orientation, and movement. SPEAR-1 differs from existing robot foundation models because it includes 3D data in its training mix.
That difference is central to the model’s purpose. By incorporating 3D information, SPEAR-1 gains an enhanced understanding of the physical world, which can make it easier to reason about how objects move through space.
Martin Vechev, a computer scientist at INSIAT and ETH Zurich, framed the issue as a mismatch between the world a robot inhabits and the knowledge inside the VLM at the core of many robotic models. “Our approach tackles the mismatch between the 3D space the robot operates in and the knowledge of the VLM that forms the core of the robotic foundation model,” Vechev says.
Open weights for embodied AI
The open-source nature of SPEAR-1 is also part of the story. In generative AI, open source language models have allowed researchers and companies to test ideas, modify systems, and move faster without relying only on closed commercial tools.
Vechev argues that robotics needs a similar path. “Open-weight models are crucial for advancing embodied AI,” Vechev told WIRED ahead of the release.
For factories and warehouses, the practical interest is clear from the source article: smarter hardware could help industrial robots manipulate objects with new dexterity. That does not mean the field has already solved general-purpose robot intelligence. It means a stronger open model gives more teams a shared base for experiments.
That can matter for startups as well as academic researchers. If teams can build on an open model, they may be able to test new hardware, tasks, and training approaches more quickly than if every project begins from a closed system or a narrow custom model.
How SPEAR-1 compares with commercial models
SPEAR-1 is described as roughly as capable as commercial foundation models designed to operate robots when measured on RoboArena. RoboArena is a benchmark that tests whether a model can get a robot to perform tasks such as squeezing a ketchup bottle, closing a drawer, and stapling pieces of paper together.
The comparison is important because the commercial race around more capable robots already has billions of dollars behind it. The source identifies well-funded startups including Skild and Generalist, as well as Physical Intelligence.
SPEAR-1 is almost as good as Pi-0.5 from Physical Intelligence, a billion-dollar startup founded by an all-star team of robotics researchers. That does not make SPEAR-1 a finished solution for every robot task. It does show that an academic open-source model can perform in the same broad conversation as commercial systems built for robot control.
The broader landscape may end up resembling the split already visible in AI software. The source article suggests that the future of robot intelligence could involve closed models from OpenAI, Google, and Anthropic, alongside open source variants such as Llama, DeepSeek, and Qwen.
The limits are still real
Even with SPEAR-1, robot intelligence is still described as being in its infancy. A model can be trained to run a robot arm so it reliably picks certain objects from a table. But real-world flexibility remains hard.
The core issue is generalization. If the robot arm changes, or the object changes, or the environment changes, the model may need to be retrained from scratch. That makes the gap between a controlled demonstration and broadly useful automation substantial.
Researchers hope that a recipe similar to the one behind large language models could eventually work for robotics: huge amounts of training data and compute. If that approach succeeds, robot models could adapt very quickly to new situations or new tasks.
The long-term implication in the source is especially significant for humanoids. Models with a general understanding of how the world works might eventually help humanoids operate in messy and unfamiliar environments. That remains a goal, not a finished reality.
What experts are watching next
Karl Pertsch, a researcher at the company Physical Intelligence, says it is too soon to know how important 3D training data will be for robotic foundation models. That caution is important: SPEAR-1 points to a promising direction, but the field has not yet settled which ingredients matter most.
At the same time, Pertsch sees the release as evidence that general robot models are moving quickly. He said, “It's really cool to see academic groups building quite general policies that can actually be evaluated across a diverse set of environments out-of-the-box, and [can] achieve non-trivial performance.” He added, “This was not possible even a year ago.”
That is the main takeaway from SPEAR-1. The model is not simply another robot control system. It is a sign that open-source embodied AI is beginning to enter territory once dominated by closed commercial efforts, with 3D reasoning becoming a serious part of the technical path forward.