This Robot Brain Thinks in 3D With Open Source Code
▼ Summary
– European researchers released SPEAR-1, an open-source AI model that serves as a brain for industrial robots to improve their grasping and manipulation abilities.
– The model was developed by INSAIT in Bulgaria to help researchers and startups build smarter hardware for factories and warehouses.
– SPEAR-1 incorporates 3D data in training, enhancing its understanding of the physical world compared to models based on 2D images.
– It performs nearly as well as commercial robot foundation models on benchmarks and is comparable to models from well-funded startups like Physical Intelligence.
– Robot intelligence remains in early stages, requiring complete retraining when hardware, objects, or environments change.
A powerful new open-source artificial intelligence model designed to serve as a brain for industrial robots has been released by European researchers, offering enhanced dexterity for grasping and manipulating objects in real-world settings. This development promises to accelerate innovation in robotics for factory and warehouse applications.
Developed at Bulgaria’s Institute for Computer Science, Artificial Intelligence and Technology (INSAIT), the model, named SPEAR-1, provides researchers and startups with a flexible tool for building and testing smarter robotic hardware. Martin Vechev, a computer scientist affiliated with INSAIT and ETH Zurich, compared its potential impact to that of open-source language models in generative AI. He emphasized that open-weight models like SPEAR-1 are essential for making rapid progress in embodied AI systems.
What sets SPEAR-1 apart from earlier robot foundation models is its integration of 3D data during training, which significantly improves the system’s spatial reasoning. While most robotic models rely on vision language models trained primarily on labeled 2D images, SPEAR-1 bridges the gap between flat image-based knowledge and the three-dimensional environments where robots actually function. This allows the AI to better predict how objects move and interact in physical space.
In performance benchmarks such as RoboArena, which evaluates tasks like squeezing a ketchup bottle, closing a drawer, and stapling papers, SPEAR-1 matches the capabilities of leading commercial robot foundation models. It even comes close to rivaling systems like Pi-0.5 from Physical Intelligence, a high-profile startup backed by top robotics experts and significant funding.
The broader race to develop more capable robots has attracted billions in investment, with companies such as Skild, Generalist, and Physical Intelligence all pursuing general-purpose robotic systems. SPEAR-1 illustrates that both proprietary models from firms like OpenAI, Google, and Anthropic, as well as open-source alternatives such as Llama, DeepSeek, and Qwen, will play important roles in advancing robot intelligence.
Still, robotic AI remains at an early stage. While it is possible to train a model to control a specific robot arm for picking up designated objects, even small changes, such as switching to a different arm model or altering the object or environment, often require the system to be retrained completely. This highlights the ongoing challenges in creating flexible, adaptive robotic intelligence.
(Source: Wired)
