This Open Source Robot Brain Thinks in 3D (www.wired.com)

Researchers at Bulgaria’s INSAIT have released SPEAR-1, an open-weight “robot foundation model” designed to give robots a better grasp of 3D physical space. Unlike prior robot models that mainly reuse vision-language models trained on labeled 2D images, SPEAR-1 explicitly incorporates 3D data into its training mix to close the gap between a robot’s real-world 3D dynamics and the 2D-centric knowledge in VLM cores. The team says the model is intended as an open platform — akin to open-source LLMs — to let academic labs and startups iterate faster on embodied AI and robotic hardware. Technically, SPEAR-1 matches the performance of many commercial robot foundation models on RoboArena, a suite of manipulation tasks (e.g., squeezing a ketchup bottle, closing drawers, stapling paper), and is close to Pi-0.5 from Physical Intelligence. That suggests 3D-aware training can materially improve generalization for manipulation tasks, though researchers caution robotic intelligence is still nascent: models often need retraining for different arms, objects, or environments. SPEAR-1’s release signals a growing ecosystem of both closed and open robot brains, and supports the idea that scaling diverse 3D data and compute — much like LLM recipe — could eventually enable more adaptable, generally capable robots.