Dominique Paul | 📍 CoRL

Two months of ablating π0.5 finetunes on a real manufacturing task. The policy now succeeds 98% of the time.

We're publishing all results, all the data, and every run, including the ones that went nowhere.

What we learned:

1/ 5x more data took us from 63% to 76%. Pure scaling was the weakest lever we found.

2/ Diversity got more out of the same hours. 4h spread across five scenes beat 4h in the eval scene by 30pp.

3/ Quality did even better. One clean, slower hour on top of the 21h: 76% → 90%. That single hour added more than the 17 hours before it.

4/ So we tried quality on its own: 1h of clean data plus 240 rollouts with human interventions. 1.7h in total, 28% → 88%. That's 12 points above the 21h model with a twelfth of the data.

5/ Inference settings add more on top, with no retraining. They took the 21h model from 76% to 93%, and the 21h+1h model from 90% to 98%.

Just as useful: what didn't matter. Learning rate, batch size, relative vs. absolute joints, image augmentation. All less important than we expected.

Full write-up linked below.

I'm surprised that several oral talks by Chinese researchers at CoRL are being presented by others because the original authors faced issues getting a US visa.

This is a very creative & interactive demo

On my way to CoRL and the British Airways flight from London to Austin has free Starlink in economy. Never expected to see so many screens with Claude Code on a flight.

We just spent two days at a company that will most likely become our next paid pilot.

At the last company, we had their parts in our lab before the visit. This time we went from first data collection to a live demo in under 24 hours.

Last time I thought six people at a demo was a lot. This morning 13 joined. The demo went well. The tasks were passing cables between machines and pressing buttons. We expected cable insertion to be harder because the plug gets occluded as it goes in, but around 200 demonstrations were enough.

Still plenty to improve:

> Setup takes too long. @Janne Stiefel has improved our setup for positioning the robots in front of the machines, but we need to get to the point where we roll the setup out of a car and plug it in.

> On one task, half the demonstrations had corrupted follower-arm states. We only noticed halfway through and had to reconstruct them with a linear model. Fine for a demo, but it can't happen again, so we added checks that flag it during collection.

> For the second time, internet speed was the bottleneck. My mobile hotspot was fast, and we moved data between the workstations and my laptop on an SSD (bringing one was a lesson from last time). Starlink doesn’t work well indoors and is limited on upload speed even when the connection is good.

> Alternating between data collection and retraining is the fastest way to improve the model, but each round still takes a few hours, even on 8x H100. Still thinking about how to improve those.

Can’t share anything from the factory due to confidentiality unfortunately but here‘s a picture of @aurel_arnold and me taking turns catching up on sleep on the way back.