daxiaorobotics/ace-brain-0.5

ACE-Brain-0.5 is a research project for a single AI model meant to act as the brain for physical robots. Rather than building separate models for separate robot skills, the team behind it built one model that tries to handle the whole loop a robot needs: sensing its surroundings, deciding what to do to reach a goal, actually acting through a robot body, checking whether the action is working, and learning from what happened to do better next time. The model organizes robot intelligence into five connected abilities: understanding space around it, making decisions, interacting with the physical world, monitoring its own progress, and improving itself over time. A single model with about 8 billion parameters is meant to cover all of this, including recognizing objects and what can be done with them, reasoning about 3D space from the robot's own point of view, planning multi-step tasks, generating the actual movements needed for navigation and manipulation, and estimating how far along a task is. The README describes a shared core that reads in different kinds of inputs and keeps a single internal picture of the scene and the task, with separate output paths that turn that shared picture into specific results such as pointing at objects, planning the next subgoal, generating movement commands, or estimating task progress. Training follows a staged approach the authors call SSR+, which builds on an earlier spatial reasoning model, specializes it for different robot skills, merges the specialized versions back together so they do not interfere with each other, and finishes with a lightweight stage that makes the outputs consistent across all the different skills. The project is evaluated on how well one model can do many different things at once, rather than on beating specialized models at a single task. The authors report that it keeps strong spatial reasoning while adding planning, executable navigation and manipulation actions, progress tracking, and the ability to improve from feedback on its own attempts. The README links a technical report and a hosted model checkpoint, along with a citation entry for the paper, but does not include installation instructions, example code, or a license.