What is tampire?

TAMPire is a hackathon project that teaches robots how to plan and carry out tasks on a tabletop, like moving a block into a bowl, starting from just a photo and a plain-language instruction such as put the red block in the bowl. It was built for a hackathon around the Gemma 4 language model running on Cerebras hardware, and its core idea is that this hardware is fast enough to make a normally slow technique practical in real time. The system works in stages. First it looks at the image and figures out what objects are in the scene and how they relate to each other. Then it turns that understanding into a formal plan made of simple actions, like picking up an object and placing it somewhere. Before the robot acts, a lightweight checker tests whether each step in the plan is physically possible, such as whether an object is blocked by something on top of it. If a step fails that check, a group of AI agents debate the failure and propose a fix, and the plan is repaired and rechecked. The project calls this group a debate council, and it is designed so a human watching a demo can see the plan fail and then watch the agents fix it live. The README describes several levels of testing, ranging from a simple simulated evaluation with no physics, up to a version using a real physics engine called MuJoCo, and further tiers that connect to established robotics benchmarks including robosuite with a real robotic arm design and an NVIDIA benchmark called RoboLab, though the most advanced levels require specific hardware such as an NVIDIA GPU running Linux that is not available on the developer's machine. The project also includes its own multi-agent vision system that estimates the 3D position of objects by combining views from several camera angles, which the README reports as more accurate than judging position from a single image. Overall this is an experimental, code-heavy research prototype rather than a polished tool, aimed at people working in robotics or AI research who want to explore whether fast language model reasoning can make automated task planning and error correction practical for physical robots.