UC Berkeley's Ken Goldberg argues 10,000 hours of open-source data is insufficient to train robot foundation models · Digg

/Tech7h ago

UC Berkeley's Ken Goldberg argues 10,000 hours of open-source data is insufficient to train robot foundation models

Story Overview

Aggregated open-source robotics datasets have hit roughly 10,000 hours across efforts like AgiBotWorld, Open X-Embodiment, and ABC, prompting claims that this scale now lets anyone train a working robot foundation model. Ken Goldberg at UC Berkeley pushes back, pointing to the practical headaches of blending recordings from robots with mismatched arms, grippers, and sensors.

6555157.5K

Original post

Ken Goldberg@Ken_Goldberg#508inTech

Chris, 10k hours (approx 1 year) is roughly 0.001 % of internet scale data. Plus it’s not clear how to merge data from different embodiments.

Chris Paxton@chris_j_paxton

I think this means you can collect ~10k hours just from open source datasets, which means basically anyone should be able to build a decent robot foundation model:

500 from the new BitRobot dataset 500 from galaxea https://huggingface.co/datasets/OpenGalaxea/Galaxea-Open-World-Dataset 3000 from agibot https://huggingface.co/datasets/agibot-world/AgiBotWorld2026 ~3000 from Open X embodiment (though it's mostly pretty bad data) https://robotics-transformer-x.github.io/ 830 from EgoDex https://github.com/apple/ml-egodex ~30 from humanoid-everyday (but good quality) https://huggingface.co/datasets/USC-PSI-Lab/humanoid-everyday 3500 from ABC https://huggingface.co/datasets/XDOF/ABC-130k

2:04 AM · Jun 25, 2026 · 6.4K Views

Open Question

Merging hardware quirks remains the sticking point

Goldberg’s objection centers on embodiment differences rather than raw volume alone, yet no detailed benchmarks or failure cases from his side have surfaced so far.

Developer Impact

Next moves hinge on cross-platform fixes

Without clearer evidence on how much extra data or new techniques would close the gap, teams eyeing these datasets for foundation models face an unresolved transfer hurdle.

Sentiment

Users are optimistic that 10k hours of open source robot datasets can support foundation models, agreeing that separate world models per embodiment plus cross-embodiment conversion will make the data sufficient.

Pos

100.0%

Neg

0.0%

3 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Related links

GitHub - apple/ml-egodex: EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video

GITHUBVia

agibot-world/AgiBotWorld2026 · Datasets at Hugging Face

Posts from X

Most Activity

VIEWS111LIKES1REPLIES1

Vai Viswanathan@vai_viswanathan

@Ken_Goldberg Any thoughts on promising directions for cross-embodiment transfer?

I’ve been looking into Latent Action Models, and it seems like there’s a lot of room for research in that space.

4h1111

RETWEETS1

Chetan@chetan_

If you include UMI style datasets there is even more.

iirc pi0 was trained on ~10k hours so definitely sizable enough to do some interesting work!

Chris Paxton@chris_j_paxton

I think this means you can collect ~10k hours just from open source datasets, which means basically anyone should be able to build a decent robot foundation model:

500 from the new BitRobot dataset 500 from galaxea https://huggingface.co/datasets/OpenGalaxea/Galaxea-Open-World-Dataset 3000 from agibot https://huggingface.co/datasets/agibot-world/AgiBotWorld2026 ~3000 from Open X embodiment (though it's mostly pretty bad data) https://robotics-transformer-x.github.io/ 830 from EgoDex https://github.com/apple/ml-egodex ~30 from humanoid-everyday (but good quality) https://huggingface.co/datasets/USC-PSI-Lab/humanoid-everyday 3500 from ABC https://huggingface.co/datasets/XDOF/ABC-130k

2h1.3K84

zeke@zshobbs

@Ken_Goldberg Look at the GR00T N1 paper from Nvidia they trained on lots of different embodiments.

https://arxiv.org/abs/2503.14734

3h52

clankr@clankrmedia

@Ken_Goldberg Definitely. A grounded engine should convert data for cross-embodiment. If this problem is solved, 10k hours of data is enough for a foundation model. Interesting work is happening there; it is better to watch them.

2h31

Michael Cho - Rbt/Acc@micoolcho

@Ken_Goldberg Yes teleop data is tough to collect; expect more to come but grow only at linear scale.

Human data (eg. ego videos, wrist cameras, tactile gloves, etc) can scale way more though.

1h22

Скандинавское математическое нытье 🍀@V4ldeLund

@vai_viswanathan @Ken_Goldberg I think cosmos just straight away threw away real actions, trained on pseudo delta actions and then post train on Droid

3h21

Justin Chen@ch3njus

@Ken_Goldberg Cross embodiment research generally seems to be picking up pace. Recent work from Karen Liu on masking out grippers in egocentric to retarget seems promising and recent humanoid wb retargeting seem pretty decent on demos.

2h13

Taro Bushidō@techietaro

@Ken_Goldberg Exactly. Rather than merging raw trajectories across embodiments, train each within its own world model first. Then combine generalizable features at a higher abstraction layer.

5h1