Georgia Tech prof. @animesh_garg predicts the next 2 years of robotics will be "the return of the RL" as the field shifts from collecting data to learning from it: "Pillar one: how do you represent the environment? Should the input be objects? Should the input be completely latent? Should the input be somehow processed with a VLM or LLM? There is debate in the field." "Pillar two is fundamentally what people now call models of the world. World models or critic models or reward models, some sort of world models that allow a very good prior on what to do." "Robotics is classically a reinforcement learning problem, so what we call post-training or recursive improvement or continued learning, those are exactly the problems that will show up in robotics in maybe a year or a year and a half." "In the same way we saw the jump from early GPTs to GPT-3 and ChatGPT, but ChatGPT by itself did not end up becoming enterprise useful. It took another two or three years from there to becoming the Claude moment." "The next two years of robotics will basically be, now that we have the data, what do we do in RL? As a lot of people call it fondly, the return of the RL." @GeorgiaTech