Joel Jang explains world action modelsA world action model connects learned representations of a physical environment with actions that can guide an embodied system such as a robot. as a way to transfer physical knowledge learned from video into robot behavior. He contrasts these priors with approaches centered on language and describes the practical friction of gathering demonstrationsImitation learning trains an agent to perform a task by learning from demonstrations of behavior rather than relying only on an explicitly written reward function., training and hosting a model for each task.
Joel Jang argues that more general software could make cheaper hardware useful to more developers. Sophia Dew presses him on actual consumer demand and use cases. He acknowledges that the released model is not best on every task and that launch activity does not yet establish a mature consumer market.
Joel Jang describes public access as a way to learn which problems users want robots to solve, rather than assuming a single winning application in advance. His predictions about consumer adoption and mobile manipulationMobile manipulation combines a robot's ability to move through an environment with its ability to handle or interact with objects. remain forecasts, not independently validated product performance.
Watch on YouTube
