Another Chinese embodied AI startup worth watching: BeingBeyond.
@ErenChenAI11K followers
27,905 views
21 Comments
Comments from the original post on X (Twitter)

@ErenChenAI
Its training data reportedly scaled from 3,000 hours with Being-H0 → 35,000 hours across 30+ robot embodiments with H0.5 → 200,000 hours with H0.7 → 500,000+ hours with H0.8, which also introduces tactile sensing. The company is also pushing hard on deployment: its lightweight Being-H-Flash model reportedly runs at nearly 20 FPS on an NVIDIA Orin NX, while its hardware ecosystem includes dexterous-hand data collection, desktop manipulation systems, teleoperation, and partnerships with robot OEMs. The founding team comes largely from Peking University, Renmin University and Zhipu AI, with PKU professor and Zhipu AI researcher Lu Zongqing as founder. China is starting to produce a distinct group of companies betting that the biggest opportunity in humanoids may be the brain layer across many different bodies, rather than building yet another robot body.
6 likes

@YuXiang_IRVL
There is no head camera for these data. That could be a problem
4 likes

@ErenChenAI
It should be compatible with other ego devices
2 likes

@robotwinnem
while normally comprised of other open datasets such as egodex and pnd adam, they also upload their training corpus onto huggingface. pretty cool!
1 like

@antopatrex1
the pure software play is a bet that you can build a robot brain without real world deployment data. genuinely curious how that works when your policy has never touched a cardboard box that was taped wrong or a shelf that's 2cm off from the CAD model.
1 like

@cogentia
What does it solve? Why don't just use tactile, (+ add modality here) gloves? Once the data is captures you can transform it into any other neural network/function approximated model I don't see any benefit other than protecting the wearer @grok truth?

@abhishekX697
UMI final boss

@SAldwais
@grok $?

@GeekParkHQ
200k hours of human video for Being-H0.7, Lu Zongqing out of BAAI and Peking, Lenovo and Zhipu in. Betting on the brain and skipping hardware entirely is the opposite of Unitree's playbook. How do you validate sim2real owning zero robots?

@YieldFarmer
CRAZYY man

@DanielSmidstrup
Pure robot brains without hardware is a bold bet, curious to see how far it goes :D

@gerardsans
@Em_Nomadic

@_Picokel
Wow

@AmyNoteApp
Unbundling the brain is clean until you realize whoever owns the body owns the deployment-failure data — the taped-wrong box, the shelf 2cm off the CAD. Human video gives you priors; the OEMs "bleeding money on bodies" quietly capture the closed-loop data you can't scrape.

@GeorgePartt
unbundling the robot brain from the hardware is the smartest play right now. let the OEMs bleed money on bodies while these guys just scale the software across everything.

@LibertyForgeMfg
The bind with brain-only is data. Embodied models need fleets running in the real world, and fleets come from whoever ships the hardware. Stay pure and you are renting your data source from the people best placed to build the brain themselves.

@stevencheng
Which foundation models are they building on?

@liquidated0x
built something similar but for trading bots, china's moving fast tho fr

@GrnEnergyBoost
My favourite part is the pure software approach.. Incredible that one model layer can scale across industries without building factories..

@nikhilkr
China really is shipping amazingly in Robotics and AI lately, and it's the fact that they genuinely understand the problem statement first.
