OpenEMMA

Why we included this project

OpenEMMA is a working, open implementation of Waymo's EMMA approach, for teams that want to experiment with end-to-end motion planning without hand-assembling a separate perception stack. The pipeline feeds front-view camera frames and text prompts through a pretrained vision-language model such as GPT-4 or LLaVA, then outputs future ego waypoints along with a written rationale for each decision. Everything packages into an installable CLI that runs against the nuScenes dataset and renders annotated trajectory overlays and video, so you can watch the predicted path on real driving scenes instead of staring at numbers. Because it leans on general-purpose multimodal knowledge plus an automatic critical-object detector, it is also a convenient testbed for how well that pretrained transfer holds up against dedicated driving stacks.