We might be solving the wrong problem in robotics. That’s what this makes clear. UMI → Universal Manipulation Interface A simple $400 gripper that lets you teach robots by demonstration. You hold it like a tool. Show the task. The robot learns. No teleoperation. No expensive hardware. No robot-specific... data. Stanford open-sourced everything → hardware, code, datasets. What stands out to me is the bottleneck. Not algorithms. Data. Teleoperation → ~35 demos/hour UMI → ~111 demos/hour And the data transfers across robots → UR5, Franka, others. The design is surprisingly practical: → GoPro fisheye lens (155° FOV) + mirrors for depth → SLAM + IMU for precise 6DoF tracking → latency matching for dynamic tasks → diffusion policies for multimodal actions Then it scales. Cheng Chi takes this further with Sunday Robotics (with Tony Zhao). A $200 glove → deployed in 500+ homes → ~10 million real-world interactions. Not lab data. Real human behavior. Their robot learns dishes, laundry, espresso → with zero robot-specific data. This is where the shift becomes obvious. From training robots in controlled environments → to learning directly from humans at scale So here’s the real question: Will robotics be unlocked by better models… or by unlocking data? #ArtificialIntelligence #Robotics #AI #Innovation #FutureOfWorkshow more

just use your damn hands!

Wow this illustrates the importance of dexterity, these grippers seem super limiting. The video is clearly a toy example, id love to see any human try and clean my dishes with those grippers. Fable has it the sink ran for a 100 years and the dish was yet to be cleaned.

This is teleops ala imitation learning. Get your facts right

a 400 dollar gripper and a phone camera doing what a 50k teleop rig used to do. the moment robotics stops being a phd thing and starts being a shop class thing is way closer than people think

this is a good solution, but it feels like an unnecessary middle man. The AI will probably just learn strait from videos or demonstrations.

JUST USE A FUCKING DISHWASHER

the missus will be pleased to experiment with you two @CaptVenk 😅

JUST LOAD THE FUCKING DISHWASHER, IT DOESN'T HAVE TO BE THIS COMPLICATED.

the bottleneck isn’t better models, it’s scalable data. Teleop locks you into slow, expensive, lab-only collection with embodiment mismatch. The q is how much IK tracking limits dynamic tasks in real deployments

Really cool! But bro is definitely using a sponge for the first time in his life. Don't teach it to robots 😅

And stop writing these bullshit lame AI generated brown mass!

That's not smart, teach them about their grippers/hands, teach them the controls, fine controls, limitations. Allow them to write and carry out their own checks and balances. Train them on it. Repeat. Use a human equivalent as the discriminator and the control. Now loop. Congrats, you now have an intelligent robotic AI to START learning by doing the very task you had a human in there teaching it. Human removed from the loop = profit = scale.

Do you have the link for the code/3d models/...?

The data bottleneck was always there.

completely agree. You can tell which of these physical AI companies truly understand manufacturing lean and legacy and those who don't.

i love how hands on this is. seeing people teach robots using simple tools like those orange grippers makes it feel way more real for small teams.

Universal Manipulation Interface (UMI) is a simple $400 gripper that lets you teach robots by demonstration. You hold it like a tool. Show it the task you want to teach. And the robot learns. The next billion dollars is in tacit training.

The problem is in real world there are unpredictable events and interruptions. You might be washing dishes and your kid run and bump into you, a human can recovers from that, can this method collects useful data to allow the robot to do the same?

The gap is not teaching robots but enabling robots to generalize from learning examples.

Sorry but I can't really do that many tasks with only 2 fingers. You're making the problem way harder than it needs to be if you design robots like this. You're also vastly underestimating how strong human finger strength and grip strength is.

i wonder ….

data unlocks scale, better models need good data

dude

This ^^

@grok find me the link to this Stanford open-sourced everything → hardware, code, datasets. Cheng Chi takes this further with Sunday Robotics (with Tony Zhao)

It’s a nice idea and experiment. But I assume nobody will actually expect to train a general robot based on like thousands of people cleaning millions of dishes? They will just decode the movement and some data and then write the intelligence that does that?

this is a good idea

UMI is a good data modality but I think the current implementations require a geometrically identical UMI gripper attached to the robot which somewhat limits its deployment possibilities (maybe this has changed in recent work?) A more fundamental issue is humans are more dynamic than most embodiments and it is unclear what to do when the humans accelerate more quickly than the hardware can handle.

can you clarify what you mean? i've been thinking about this same problem, but i can't put it into words. is the bottleneck training data? (but we have simulation). is the problem the sim to real gap? (but again, Nvidia has this kinda answered). what are we doing wrong?

Data's the unlock. 3x demos/hr is huge for iteration speed.

Yeah but we’re pretty terrible at a lot of things.

Training our replacements faster than ever.

hhahha, this looks good, but will it be genuinely useful

Or you deploy smart glasses with the right sensors

pretty good. although the two-prong gripper used here is far from ideal.