WITH this week introduced a brand new model for training robots. Instead of the usual set of focused data used to learn recent tasks for robots, the tactic goes far and mimics the vast repositories of data used to coach large language models (LLMs).
The researchers note that imitation learning – where the agent learns by following an individual as they perform a task – can fail when small challenges are introduced. This may very well be things like lighting, a distinct environment or recent obstacles. In such scenarios, the robots simply don't have enough data to attract on to adapt.
The team used models like GPT-4 as a sort of brute force data approach to problem solving.
“In the language domain, the info are all just sentences,” says Lirui Wang, lead writer of the brand new paper. “In robotics, given the heterogeneity of the info, we want a distinct architecture if you wish to pre-train in the same way.”
The team introduced a brand new architecture called Heterogeneous Pretrained Transformers (HPT), which brings together information from different sensors and different environments. A transformer was then used to merge the info into training models. The larger the transformer, the higher the performance.
Users then enter the robot design, configuration, and desired task.
“Our dream is to have a universal robot brain which you could download and use on your robot with none training,” David Held, an associate professor at CMU, said of the research. “Although we’re only within the early stages, we’ll proceed to work hard and hope that scaling results in a breakthrough in robotics policy, as has been the case with large language models.”
The research was initiated partially by the Toyota Research Institute. Last yr at TechCrunch Disrupt, TRI presented a technique for training robots overnight. Recently, the corporate entered right into a groundbreaking partnership that can mix its robot learning research with Boston Dynamics hardware.
