This Embodied LLM is...

PaLM-E is a new LLM from Google that is both embodied in multimodal. Excitingly, it shows positive transfer across different robotics tasks. While the prospects are exciting, it is unclear what other conclusions can be drawn from the work.

Outline
0:00 - Intro
1:34 - How It Works
6:47 - Robotics Tasks
11:25 - Results
26:50 - Takeaways

Social Media

Sources:

[earleyelisha]

Good review and your critique was spot on. Hype and unclear claims from research groups aren’t helpful for progress.

[davidyang]

would your embodied model be n-shot learning? you get some interactions with environment to calibrate yourself. does this go under meta learning?

[chrisliu]

Small nitpick: I think by 4B they mean param count?

[zack]

Am I off with the impression that the paper seemed written more for business execs / investors than fellow researchers?

[IgarokSpider-htve]

Can you review most accurate partly supervised models for calculations on encrypted data, and if there are good pre-trained ones

[Andrea]

Can anyone explain how, despite the robotic data being just a small fraction of the mixture, the model manages to extract from it enough knowledge to perform well on the robotic benchmarks? I would guess they could have achieved better performances by sampling more frequently from the 3 robotic datasets they have.

[wpgg]

As always, really cool. Thanks for that !
But to be honest, it's kinda strange anime...

[chillmathematician]

Hi, Eden Meyer, can I know how reinforcement learning researcher learn and do RL analysis? I already go though most of the model free RL, and now I am looking into learn about RL Theory to improve my fundamental knowledge of RL.