Last week I presented real-time chunking (RTC) at NeurIPS, and we did a live coffee demo the very same evening. To celebrate, we're releasing a (very short) follow-up paper describing a training-time variant of RTC, which is what we've actually been using in our demos!

The method is stupidly simple -- we simulate delay at training time by conditioning on action prefixes. It only takes about 8 lines of code to implement, but it works just as well as inference-time RTC without the extra computational overhead.

Check out the paper here: And the code: Hopefully some people find this useful!

Coffee making is the current standard.

Incredibly cool! Did you 3D print those arms in house? 🦾

cool

Do you think generalist is doing something very different on the training front - besides their umi setup?

Amazing! Absolutely stunning!

first robot actually moving in a way i get that its the endgame

Nice work! Do you think it will replace the more complex RTC alternatives in the near future?

it all depends on your constraints. inference-time RTC is still more convenient. however, we already do a lot of post-training so we may as well add something there, and this simple method seems to work well enough. I'm sure ppl will find other methods that work better.

Thanks! Curious to see what other methods emerge.

Interesting approach with the training-time variant. Curious how it compares to existing models in terms of performance and versatility. Got any benchmarks?

beautiful!