Noematrix

At Noematrix, an embodied-intelligence company building vision-language-action models, I post-trained FastWAM, a pretrained world-action model, for real-robot pick-and-place tasks such as placing eggs into trays and returning tools to toolboxes.

I deployed the post-trained checkpoint to a dual-arm real robot through the team's inference server, integrating it with the gripper and device pipeline. Real-robot pick-and-place success rose from 0% with the pretrained baseline to 33%. The evaluation work here is what motivated robostats.