This Open Source Robot Mind Processes Information in 3D

Today, European roboticists unveiled an advanced open-source AI model designed to serve as the cognitive engine for industrial robots, enabling them to handle and manipulate objects with newfound skill.

The model, named SPEAR-1, was created by researchers from the Institute for Computer Science, Artificial Intelligence and Technology (INSAIT) in Bulgaria. It has the potential to assist other researchers and startups in developing and testing more intelligent hardware for factories and warehouses.

Similar to how open-source language models have allowed researchers and companies to explore generative AI, Martin Vechev, a computer scientist at INSIAT and ETH Zurich, believes that SPEAR-1 can facilitate rapid experimentation and iteration for roboticists. “Open-weight models are pivotal for progressing embodied AI,” Vechev told WIRED prior to the release.

SPEAR-1 stands out from current robot foundation models by integrating 3D data into its training process. This allows the model to better comprehend the physical world, making it easier to understand the movement of objects in physical space.

Typically, robot foundation models are built upon vision language models (VLMs), which possess a wide but often limited understanding of the physical environment due to training primarily on labeled 2D images. “Our method addresses the disconnect between the 3D environment in which the robot operates and the knowledge base of the VLM at the core of the robotic foundation model,” Vechev explains.

When evaluated on RoboArena—a benchmark that assesses a model’s capability to command a robot in tasks like squeezing a ketchup bottle, shutting a drawer, and stapling sheets of paper—SPEAR-1’s abilities are comparable to those of commercial foundation models designed for robot operation.

The drive to enhance robot intelligence is already backed by billions in funding. The commercial promise of versatile robots has led to well-funded startups, including Skild, Generalist, and Physical Intelligence. SPEAR-1 performs nearly as well as Pi-0.5 from Physical Intelligence, a billion-dollar startup formed by a distinguished team of robotics experts.

SPEAR-1 indicates that the journey toward developing smarter robots may require both proprietary models from companies like OpenAI, Google, and Anthropic, along with open-source alternatives such as Llama, DeepSeek, and Qwen.

However, robot intelligence is still very much in its early stages. While it is achievable to train an AI model to operate a robotic arm for consistently picking specific objects from a table, in practice, the model would necessitate complete retraining if a different type of robotic arm is used or if the object or environment changes.