New models make robotics simulation more scalable

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 10:48

   if robot labor was too cheap to meter. What's unique about today in 2026, we're almost at the end of the year, that makes all of this possible? And why couldn't this have happened five years ago? It's a good question. So one thing is we have the transformer, right, which is pretty magical. And the transformer has given us kind of two big things that are really helpful in this moment of robotics. One is just LLMs. And LLMs kind of are this incredible model that can suck up all this understanding about the world. And you can use it in a bunch of ways to power robots. And we see this in kind of leading VLAs, use pre-trained LLMs as kind of a backbone. We're also seeing new approaches for RL post-training using directly frontier LLMs. So that's one difference from transformers. The other difference is like we

2. 11:37

   actually can do robot-specific models now, right? We have action chunking diffusion transformer heads that you can kind of stack on to an LLM and you can generate a robot motion through space, right? We started this company in 2021. We were using tutors to teleoperate robots and label data for convolutional neural networks. And there was just a limit to how expressive that could be and what jobs or robots could do.

3. 11:58

   What else about regulation? or factories, or even these customers you're talking to being more open to having robots within their assembly lines is different today? Yeah, I think people are talking a lot about re-industrialization today, but these aren't one-year trends. I think they're not even five or 10-year trends. These are very, very long-term secular trends, right? Of how is the United States and China building our industrial capacity? What is shifting labor forces look like?

4. 12:25

   And we saw this even when we started the business in 2021 with COVID. I think that was kind of a sign of what was to come. And we're seeing that sort of continue of like these businesses in the physical economy, we strictly need them. We want them to be better. We want them to make more stuff and to make stuff for cheaper, but they don't pencil out under the current paradigm. So we just need robots. Yeah. We need robots. People are saying GPT Astra, GPT-6 is like this GPT-3 moment, but for robotics. Do you feel that as well? Yes. How come? It's a very good model. It understands spatial reasoning in a really special way that we haven't really seen with prior models.

5. 13:04

   By the way, this is also true of Opus 5.5. I think all the big labs are post-trading on the shape of data now. And I think it's going to unlock two things. One is, I think we've already seen in robotics, you get a better backbone for your robot-specific model, and your robots get better, and they get kind of more general world knowledge for free. And that's great. It benefits everybody. I think the other thing, which is really unique and new that we're doing a lot of, is Astra can build simulation environments very quickly that mimic what the robot actually experiences. And simulation used to be something where to do simulation well in robotics, you need teams of simulation engineers who are maintaining simulation environments. That wasn't something that could scale, even though simulation theory is like this very plentiful thing. It's just

6. 13:44

   software. And now simulation actually can scale. And we can use that as another tool in our toolbox in robotics to build data distributions to train our models from. Is simulation good enough yet? Yes. Simulation is very, very, very good. There's a whole shape of problems involving rigid body manipulation where simulation can get you very close to good. You still need a robot, right? You need to be able to profile the real world.

7. 14:07

   You need to be able to measure the real world. You need to be able to know that your simulation matches what the real world is. And then you need to be able to close the sim to real gap. So all of these things eventually need a robot. And obviously to actually create value in the real world with a robot, you do need a robot. But simulation helps massively. So then where does Tudor come in here?