Although there is a lot to do, Tedrake claims that all the evidence existing suggests that the approaches to LLM also work for robots. “I think it changes it all,” he says.
Proceedings in robotics have recently become more challenging, of course, and video clips showing commercial humanoids perform elaborate duties such as Loading refrigerators Or Dulling of garbage seemingly ease. YouTube clips, however, can be disingenuous, and humanoid works are usually either teleoperative, carefully programmed or trained to perform one task in very controlled conditions.
The novel work of Atlas is a massive sign that robots are starting to experience the type of equivalent progress in robotics, which eventually led to general models of languages that gave us chatgpt in the field of generative artificial intelligence. Ultimately, such progress can give us robots that are able to easily operate in a wide range of messy environments and are able to quickly learn novel skills – from welding pipes to Espressos – without broad retraining.
“It’s definitely a step forward,” says Ken Goldberg, a robotic from UC Berkeley, who receives funds from TRI, but was not involved in the work of Atlas. “Coordination of legs and arms is a great deal.”
Goldberg says, however, that the idea of emerging behavior of robots should be treated carefully. Like the surprising abilities of vast language models, they can sometimes be traced on the examples contained in their training data, says that robots can demonstrate skills that seem more creative than in reality. He adds that it is helpful to learn about detailed information about how often the robot is successful and how he fails during experiments. TRI was previously lucid in work done by LBMS and can issue more data about the novel model.
Whether the straightforward scaling of data used to train the robot models unlock more and more emerging behaviors remains an open question. On debate It takes place in May at the International Robotics and Automation Conference in Atlanta, Goldberg and others warned that engineering methods also play an vital role.
Tedrake is convinced that the robotics are approaching the inflection point-such, which will allow more real employ of humanoids and other robots. “I think we have to take these robots out of the world and start doing a real job,” he says.
What do you think about novel Atlas skills? And do you think we are going to a break in the style of chatgpt in robotics? Let me know your thoughts at ailab@wired.com.
This is the edition Will Knight’s AI Lab newsletter. Read previous newsletters Here.