European Roboticists Today A powerful open source software has been released. artificial intelligence A model which acts as the brain for an industrial robots—helping them grasp and manipulate things with new dexterity.
This new model is a great improvement. SPEAR-1Researchers at Bulgaria’s Institute for Computer Science, Artificial Intelligence and Technology, INSAIT, developed the software. The project could be used to help startups and researchers build and test smarter warehouse and factory hardware.
Martin Vechev is a computer science professor at INSIAT Zurich and ETH Zurich. He says SPEAR-1 will help roboticists experiment quickly and improve their work. “Open-weight models are crucial for advancing embodied AI,” Vechev spoke to WIRED before the publication.
The SPEAR-1 robot differs in its use of 3D data for training from other foundation models. It gives it a better understanding of how physical objects behave in space.
In general, robot foundation models are built on top vision language model (VLM) that have a limited understanding of physical reality because the training is usually done with 2D labels. “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.
When measured, SPEAR-1 has a similar capability to commercial robot foundations. on RoboArena, a benchmark This test measures a model’s capability to make a robot do tasks like squeeze a bottle of ketchup, shut a drawer and staple paper pieces together.
Race to improve robot intelligence already underway billions of dollars The future is riding on it. Commercial potential for generally-capable robots led to well-funded startups such as Skild The following are some examples of how to get started: Generalist besides Physical Intelligence. It is about as good as Pi 0.5, from Physical Intelligence. a billion-dollar startup founded by an all-star team of robotics researchers.
SPEAR-1 suggests that in the quest for building more intelligent robots, it may be necessary to use both closed-source models, such as OpenAI’s, Google’s and Anthropic’s, as well open-source variants of Llama DeepSeek Qwen.
But robot intelligence is in its early stages. You can train an AI to control a robotic arm to reliably select certain items from a desk. The model must be trained again if the robot arm or object is changed or if there are changes to the surrounding environment.
