Stanford's Shuran Song introduces Behavior Prompting, enabling robots to learn new tasks from a single human demonstration · Digg

/TECH

Stanford's Shuran Song introduces Behavior Prompting, enabling robots to learn new tasks from a single human demonstration

SS#704

The method utilizes high data diversity instead of high repetition.

Combined views

3.6K

2 posts, first seen 4h ago

33 likes

Stanford's Shuran Song introduces Behavior Prompting, enabling robots to learn new tasks from a single human demonstration

SS#704

Source Posts on X

Cluster sources

Shuran Song@SongShuranTECH

What makes LLMs so amazing is that you can just prompt them with the right example to make them do a new task. Can robots do the same with “Behavior Prompting”? It turns out to be possible, but only with the right data. The key lesson for us: it’s not the amount of data that matters most, but the diversity of tasks. Given the same data budget, you’re better off collecting 1000 different tasks × 5 demos than 5 tasks × 1000 demos, which is very different from how we typically collect robot data today! There are more tricks for making it work. Check out @austinapatel's 🧵 to learn more!

Combined views

3.6K

2 posts, first seen 4h ago

33 likes saves

16 saves

16 reposts

Sentiment

Positive0%0%Negative

Guardrails removed spam, off-topic, unclear, or duplicate replies.

Useful Links

Behavior Prompting Policy(behavior-prompting.github.io)viaShuang Li

Digg Deeper

Ask a question below.

Published answers will appear here.