Embodied AI
Embodied AI is the part of artificial intelligence that has to work in a body: how robots see, move, grasp and learn. This section covers robot foundation models and vision-language-action models, world models, whole-body control, robot hands, training data, and the simulators where most robot skills are now learned. We read the papers, model cards and licenses, say what was shown on real hardware and what only ran in simulation, and note whether code and weights are actually released. When a lab claims a big jump, we compare it with the earlier results we have covered, so you can see whether it is new or a familiar idea with a new name.
New here? Start with what a robot foundation model is, world action models explained and how sim-to-real training works. For the data question, read why training data is the bottleneck and the Open X-Embodiment dataset explained. For the cost of training, see SuperMimic’s one-GPU humanoid result. New research and model releases are listed below.