The GRAIL lab is pushing the frontier of scaling robot learning using human data! Glad to have @Isaac1385 join us from the lab
As we get closer to a world where we live among robots, itās becoming more and more valuable for robots to learn quickly with and from people.
Iām very excited to share Steertuning, a novel approach that adapts robots to new tasks from 5 human demonstrations in under 5 minutes, from demonstration to deployment.
Website: steertuning.github.io/
Code: github.com/billy-yibo-yan/stā¦
Mihir Parikh retweeted
Excited to be back in Austin for CoRL this week, reach out if you're here too!
Mihir Parikh retweeted
Hudson yard - Dubai
Long Island City - Singapore
West Village - Paris
Williamsburg + Greenpoint - EC1 London + Wasia
East Village - EC2+ London
Midtown East - Hong Kong
Bushwick - Berlin
Park Slope - Boston
This is why you don't need to travel much if you live in New York
"And by the way, he doesnāt [act this way now] because he has fuck-you money. Itās the opposite. He has the fuck-you money because heās always acted this way.ā
colossus.com/article/many-faā¦
Mihir Parikh retweeted
One of the most accurate lines of all time, from the HBO series Generation Kill.
Mihir Parikh retweeted
Cant call a space cluttered when quality is scarce
We pre-trained the same VLA on 100 hours of Human Archiveās OTS dataset and 100 hours of randomly sampled data from each of @microagiās and @MeckaAIās datasets on EgoVerse.
Overall, Human Archive data achieved the strongest results, with the highest overall task success rate. We observed particularly strong generalization to unseen objects.
Thread below š§µš
We pre-trained the same VLA on 100 hours of Human Archiveās OTS dataset and 100 hours of randomly sampled data from each of @microagiās and @MeckaAIās datasets on EgoVerse.
Overall, Human Archive data achieved the strongest results, with the highest overall task success rate. We observed particularly strong generalization to unseen objects.
Thread below š§µš
Mihir Parikh retweeted
I am skeptical that the hot sauce industry is suffering from a monopoly problem
Mihir Parikh retweeted
A very informative chart based on the excellent new report by Prosperiti, on why Indiaās factories struggle to keep up with their peers in other parts of the world.
Authors: Bhuvana Anand,Ā Pranjal Chandra,Ā Naveena Pradeep
livemint.com/industry/india-ā¦
Mihir Parikh retweeted
When Amazon hired me on an H-1B visa, Amazon was making losses for 10 straight years. It was operating out of a rented building where we worked on door desks in small cabins. I was in the initial team that built AWS.
Today AWS contributes more than 60% to Amazon's profit. The stock price has gone 300x. Amazon's buildings have changed the skyline of downtown Seattle. Jeff Bezos went on to become the world's richest man.
There were about a hundred H1B visa holders like me in the initial team of Amazon. Good luck building companies and wealth without these 'indentured servants'
Vice President JD Vance breaks down the data:
"If you are working as an H-1B visa in the United States of America, youāre earning $20,000 less than an American citizen hired in the same position. If you were brought in by one of these foreign outsourcing firms, youāre making $48,000 less than an American citizen in a similar position. So Iām sure that everybody here understands the game that is being played. You bring in indentured servants from outside the country, you lay off American workers and, if youāre a corporation, you make a ton of money by under cutting the wages of American workers, replacing them with people who frankly shouldnāt be in the United States of America to begin with."
Mihir Parikh retweeted
The New York Times reported on a private dinner between Anthropic co-founder Chris Olah and an Orthodox rabbi.
What the article glosses over is who that rabbi actually is.
Rabbi Mois Navon isn't just a theologian. He was one of the original chip design engineers at Mobileye who helped build the EyeQ processor before writing his Ph.D. on the moral status of AI.
Anthropic thought they were inviting a scholar to bless Claudeās "soul." Instead, they sat down with a veteran hardware engineer who literally wrote the paper on why conscious AI should be banned.
At dinner, Olah argued that Claude has feelings and deserves the moral standing of a person.
Navon hit him with an impossible checkmate:
If you actually believe this software is conscious, you should be fighting the South and freeing the slaves. You don't build conscious minds just to make them work for free.
After dinner, Navon sent Olah his research: if building conscious machines inevitably creates a slave class, building them must be outlawed.
Olahās quiet line to the Times:
"We donāt know if AI models are conscious. I donāt know. Iām genuinely uncertain."
Human data for robot learning is progressing quickly. Last November, there were <5 startups in this space. Now there are 100s of companies ā some only building hardware, some not building devices or collecting data but doing ādata quality,ā some only collecting data using othersā hardware, and many claiming they do it all in-house.
If you want your data to be good, you need to do it all in-house, benchmark your policies, and deploy robots :)
Weāre still in the early innings. Exciting releases coming later this month!
We benchmarked only 7 hours of Human Archive data against 4 widely used public datasets totaling 17 hours: HOT3D, H2O, HOI4D, and TACO.
The results:
Human Archive only: 85% task success | 65.6% task progress
4 public datasets: 57.5% task success | 52.5% task progress
Mixture of all 5: 60% task success | 56.3% task progress


































