While there is a lot of work to be done, so far, all the evidence suggests that the method used for LLM is also applicable to robots, Tedrake said. “I think that’s changing everything,” he said.
Of course, measuring the progress of robotics is becoming increasingly challenging, with videotapes showing the complex chores of commercial humanoids, such as Loading refrigerator or Take out the garbage It looks relaxing. However, YouTube clips can be deceptive, and humanoid robots tend to be hand-made remotely, carefully programmed, or trained to perform a task under very controlled conditions.
The work of the new map atlas is a big signal that robots are beginning to experience the same advances in robotics, ultimately leading to the common language model chatgpt In the field of generative AI. Ultimately, such progress can allow us to run easily in a variety of messy environments and be able to quickly learn new skills, from welding pipes to making espresso without extensive retraining.
“It’s definitely a step forward,” said Ken Goldberg, a robotist at UC Berkeley. “The coordination between legs and arms is huge. ”
However, the idea of emergency robot behavior should be carefully addressed, Goldberg said. Just as the surprising ability of large language models can sometimes be traced back to the examples contained in their training data, robots may exhibit newer skills than they actually are. He added that it is helpful to know the frequency of robot success and the details of failures during the experiment. The work TRI has done on LBMS before is transparent and will likely release more data on the new model.
Whether simply extending the data used to train a robot model will unlock more and more urgent behaviors remains an open question. exist debate Goldberg was held at the International Conference on Robotics and Automation in Atlanta in May, and others warned that engineering methods will play a major role.
Tedrake firmly believes that robotics is approaching a turning point that will make human robots and other robots more realistic. “I think we need to drive these robots out of the world and start doing real work,” he said.
What do you think of Atlas’ new skills? Do you think we are going to make a Chatgpt style breakthrough in robotics? Let me know what you think on ailab@wired.com.
This is a version Will Knight’s AI Lab Communication. Read previous newsletters here.
