We think the most scalable source of robotics data is general video. A model trained to predict that video learns how objects behave, how hands move and what a task looks like partway through, giving it a head start on new embodiments.
Fine-tuned on Google’s RT-1, a robot it never saw in training, Praxis-1 reaches 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy.