superxslam/supermap

SuperMap is a research project from AirLab at Carnegie Mellon University, presented at the Robotics: Science and Systems conference in 2026. It describes itself as a living spatial memory for embodied AI: a system that lets a robot perceive its surroundings, remember how the space changes over time, and use that memory to reason about actions. Under the hood, it is a training-free system that builds a persistent map by combining fast geometric SLAM, the technique robots use to build a map while tracking their own position, with slower, open-vocabulary perception that can recognize a wide range of objects without being retrained. The result is described as a 4D scene graph, meaning the map stores information about where objects are and how their state has changed across time, not just a static snapshot. The project claims the memory supports both visual-language navigation, where a robot follows instructions given in plain language, and longer-horizon reasoning tasks. It says the system works with different object detectors such as Grounding DINO and YOLOE, can run fully offline on recorded data or live on a robot through ROS2, and has been tested continuously for two hours across the CMU campus. As of this README, the code itself has not been released yet. The authors state it will be published after the RSS conference, so right now this repository mainly serves as a project page, paper link, and demo video, plus a citation entry for anyone referencing the work. Running it, once released, is expected to need an Ubuntu machine with an NVIDIA GPU with at least 16 gigabytes of memory and Python 3.10 or newer, plus a ROS2 setup for live use on real hardware. There is a Discord server for questions in the meantime.