wtf is geometric-action-model?

GAM (Geometric Action Model) is a research project that helps robots learn to follow natural language instructions. Instead of training a robot from scratch, it takes a model that already understands 3D geometry and physical space, and adapts it into a single system that handles seeing, predicting, and deciding what to do next. The goal is to make robot training more efficient and accurate. At a high level, the system uses a pretrained "geometric foundation model" as its core engine. This means the model has already learned a general understanding of shapes and spatial relationships. The researchers then fine-tune this model so it can watch a scene, understand a verbal command like "pick up the red mug," predict what should happen next, and output the specific movements the robot arm needs to make. It packages all of this into one shared system rather than separate modules for perception, prediction, and action. This project is built for robotics researchers and engineers working on robotic manipulation tasks. A concrete use case would be a team testing how well a robot arm can perform everyday tasks in a simulated kitchen or workspace, such as opening a drawer or moving objects around based on spoken commands. The repository provides everything needed to run and evaluate this model on standard robotics benchmarks called LIBERO and LIBERO-Plus. What is notable about the project is its performance and efficiency. The released model achieves high success rates on standard benchmarks (97.6% on LIBERO) and processes its decisions quickly, taking under 7 milliseconds per decision. It is also built to scale across multiple graphics cards for training, reflecting a serious research-grade implementation rather than a simple prototype.