
Lessons from Pieter Abbeel
The items below are things Pieter Abbeel said in interviews about robotics, reinforcement learning and learning AI. — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "traditional robots are just doing repeated motion and it's great to build the cars, but it's very limiting in what they can do". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "what I'm really curious about is how we can get this notion of play into our AI agents". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "you don't know what they're going to want to do later, what tasks you want to give them later". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "we put a reward learning signal to focus it on the task or we put a play signal to make sure it collects the most interesting data possible". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "Can we just have a ton of images, pre-train on that, and then be ready to learn a classifier very quickly later". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "we are actually getting there and these robots are providing real value because we're hitting these multiple nines of reliability". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "keep questioning things". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
On the SuperDataScience podcast, Abbeel said: "I look for a lot of passion in the people that I'm interviewing and recruiting, and try to see if they really care and are excited about these subjects". — Pieter Abbeel on the SuperDataScience podcast (episode 503, September 2021), transcript.
In an ODSC interview, Abbeel said: "The easiest way to explain reinforcement learning is by first explaining supervised learning". — Pieter Abbeel, ODSC Lightning Interview on deep reinforcement learning (2022).
In an ODSC interview, Abbeel said: "in reinforcement learning, you just specify what you want and then the agent is supposed to figure out from its own experimentation what's the best way to achieve this". — Pieter Abbeel, ODSC Lightning Interview on deep reinforcement learning (2022).
In an ODSC interview, Abbeel said: "it's simpler, from a human perspective, as you don't have to exhaustively give examples". — Pieter Abbeel, ODSC Lightning Interview on deep reinforcement learning (2022).
In an ODSC interview, Abbeel said: "it can actually exceed human capabilities". — Pieter Abbeel, ODSC Lightning Interview on deep reinforcement learning (2022).
In the Heroes of Deep Learning interview, Abbeel said: "Make sure you try things yourself. So not just reading things or watching videos but try things out". — Pieter Abbeel, Heroes of Deep Learning (DeepLearning.AI), quoted advice.
On the Lex Fridman Podcast, Abbeel said: "effectively RL is teasing that apart. It's trying to say, okay, what is consistently there when you get a higher reward and what's consistently there when you get a lower reward". — Pieter Abbeel on the Lex Fridman Podcast (#10, December 2018), transcript excerpts.
On the Lex Fridman Podcast, Abbeel said: "linear feedback control is extremely successful. It can solve many, many problems surprisingly well". — Pieter Abbeel on the Lex Fridman Podcast (#10, December 2018), transcript excerpts.
On the Lex Fridman Podcast, Abbeel said: "relatively simple control architectures can already do a lot". — Pieter Abbeel on the Lex Fridman Podcast (#10, December 2018), transcript excerpts.
In a keynote on generalizable robotics, Abbeel said: "the thing that really excites me from an academic Viewpoint is this notion of generalizable learning". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).
In a keynote on generalizable robotics, Abbeel said: "predicting the next frame that's going to come in a video or a sequence of next frames requires a very deep understanding of the natural world". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).
In a keynote on generalizable robotics, Abbeel said: "the higher the entropy the more diverse parts of the environment the agent has visited the lower the entropy the less the agent has explored". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).
In a keynote on generalizable robotics, Abbeel said: "by making this small change before we run Dr Q we just do a little bit of play the learning curve shoots up a lot faster". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).
In a keynote on generalizable robotics, Abbeel said: "if you're not very precise about how you specify how things should be done things can go very wrong". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).
In a keynote on generalizable robotics, Abbeel said: "collect human preferences feedback look at pairs of trajectories say which one you prefer over which other one". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).
In a keynote on generalizable robotics, Abbeel said: "the biggest data source will be internet data just because there's so much data there that data will mostly teach our robot". — Pieter Abbeel, Paving the Path to Generalizable Robotics (keynote; auto-generated captions).