Stanford Webinar - The Frontier of Deep Learning for Robotics, Chelsea Finn

Today's robots excel at performing very specific tasks within a narrow and controlled environment. But, when faced with a novel situation, their highly specialized training doesn’t enable them to adjust on the fly. In the future, could robots learn to adapt to new tasks they haven’t been trained to do?
One possible solution to this problem is deep learning. While deep learning is expanding the capabilities of both machine learning and reinforcement learning, it also has the potential to unleash new possibilities for robotics. Join Professor Chelsea Finn in this discussion of modern deep reinforcement learning algorithms, and learn more about their usefulness towards solving ambitious challenges in robotics.
For more information about Stanford's Artificial Intelligence programs visit: stanford.io/ai
#deeplearning #robotics

[@user-uc6nv2hh1c]

Wonderful, Thank you!

[@elirothblatt5602]

Great video lecture, thank you!

[@stanfordonline]

Hi Eli! Glad you enjoyed it, thanks for watching.

[@debbyfrederickesq9486]

👍🐝

[@RalphDratman]

I am musing on why these tasks are so difficult for robots ro perform. The. most general answer, I think, is that humans are more amazing than we usually realize. Other animals -- with a few quite limited exceptions among the primates -- cannot master most of these tasks either. Another reason is that young humans receive huge amounts of expert supervision and personalized training by adults, while robots are not very amenable to such training. And note that training and supervision of young humans by older ones is only possible because the young ones are small and not very strong. If the robots could likewise start out small, weak and incidentally good at mimicry, it would be far easier for humans to train them.

[@labsanta]

My Learnings Robotics Q&A Q: What are some of the tasks that robots can perform with artificial intelligence? A: Robots can perform tasks such as slotting a battery into a remote control, tearing off a piece of tape and putting it on a box, putting a shoe on a person, and candy unwrapping. Q: What are some challenges associated with getting policies to do certain tasks? A: One challenge is being able to react quickly to what's happening, such as finding where the flap is on the candy and opening it. Another challenge is the amount of data needed to complete certain tasks. Q: What are some of the most promising approaches for building robots with true artificial intelligence? A: Scaling up data, allowing robots to adapt at test time, and embodiment are some of the most promising approaches. Q: What are some resources for learning more about robotics? A: Some resources include the Majoco physics simulator, the Robot Brains podcast, robot conferences such as Robotic Science and Systems and the Conference on Robot Learning, and course content. Q: What advice would you give to someone new to the field of robotics? A: Start by getting your feet wet, try building a robot from scratch or playing around with robots in simulation. Learn more about reinforcement learning and explore resources such as the Robot Brains podcast and robot conferences. Key Points Robots can perform a variety of tasks with artificial intelligence, from simple actions like slotting a battery to more complex actions like unwrapping candy. Reacting quickly to situations and acquiring enough data are some of the challenges in getting policies to do certain tasks. Scaling up data, allowing robots to adapt at test time, and embodiment are promising approaches to building robots with true artificial intelligence. Resources for learning more about robotics include simulators, podcasts, conferences, and course content. To get started in robotics, try building a robot from scratch or experimenting with simulations, and explore resources like the Robot Brains podcast and robot conferences.