Roberto

Anyone who is looking for Assembly as a service done by Robotic cells? Let me know. DMs are open. (currently based in germany. Nextis offers Aaas worldwide)

@wagecattle @christianhdkwls 👀👀

sign up today!

@techietaro Thanks for the advice. Will do some A/B soon. :)

@robertorobotics I'd freeze the perception stack and iterate only on the manipulation head. Should cut cycle time without regressing that 60% progress.

With my custom architecture I now achieve on average 60% progress on my gearbox assembly task. The original average progress of pi0.5 base (after 10k fine tuning) was 14% . Trajectory of progress is looking good. Now have to reduce the cycle time of iterating on a base policy.

@jaikr_x Cool! Yeah i designed them

@robertorobotics I like those grippers. They remind me of what Dyna have. Did you design them?

@BrangstrupL @vai_viswanathan no so far only supervised BC, with some smarter conditioning and sampling.

@robertorobotics @vai_viswanathan Are you doing any offline or online RL?

@wagecattle @karpathy Also do Elon

@karpathy now ranks #1 in the leaderboard

@DominiqueCAPaul lol

@robertorobotics Thought in my head "agree" and then noticed you were talking about robots not humans lol.

everyone benchmarks success rate. the metric that actually determines economic usefulness is TTLNT -> time to learn new task. and the real unlock is when the policy is substrate-independent: doesn't care which cell it runs on.

@DominiqueCAPaul My custom build. They are on my github if you are interested: github.com/robertorobotics

@robertorobotics What kind of arms are those?

@RohanSeelan Thanks :) 100 episodes!

@robertorobotics This is awesome! How many episodes were used to train the base policy?

@vai_viswanathan Mainly intervention data, with advantage labels and additional stage labels etc. Intervention data is most important.

@robertorobotics super cool. what type of architecture are you using? is there any pretraining or language conditioning?

@HyperLogistix @enactic_ai Thanks! will look into that.

@enactic_ai @robertorobotics

After a long wait since our last announcement, OpenArm 2.0 is finally here. We're expanding from a robotic arm into a standard evaluation environment for Physical AI research, anchored by OpenArm Cell. - OpenArm Cell for reproducible eval - New pinch-type end effector - Standardized cameras - Redesigned J5 wrist (natural teleop) - VR teleop - Long-term stable release github.com/enactic/openarm

@SGhashghaei @speedrun Good Luck!!

@imherefortheoil I have the same plan. Yeah i think context is a good guess. It is getting to computationally heavy.

@robertorobotics $100/month Context might be getting too long. Last prompt ran for 62 minutes

Happening to anyone else?

Life is electric. Get after it.

@Enrico23850322 Legend

Just taking matters into my own hands, inspired by my Bro to do it. @robertorobotics

Spun up an auto-labeler for my robot intervention datasets. Labels task stages, aligns negative/correction spans around human takeovers, and drops everything into a reviewable video timeline before touching the dataset. Manual labeling was becoming the bottleneck. This should 10x the iteration loop.

every time in am in awe when these two go to work. Currently codex is employee of the month at Nextis but cannot wait for them to get the title.