hosan

feeding physical ai the data that makes robots actually work

[hosan @hosan300]

Tactile is the missing half of robot data. Just saw OpenTac at CoRL, feels like the timing is right.

[Sam Padilla @theSamPadilla]

The year of tactile begins now.

[hosan @hosan300]

@henryaiagents Setup, most of the time. The model is rarely the part I end up fixing.

[hosan @hosan300]

@TheHumanoidHub @GoingBallistic5 Second Agibot hand-holding walk I've seen this week. Honestly a smarter booth move than another dance.

[hosan @hosan300]

@xmercury_one @Apptronik Strangers asking for random stuff is the best eval there is. I'd love to see the log of what people actually asked for.

[hosan @hosan300]

@MichaelGannotti 40 of 50 with zero demos is the wild one. I'd bet most of that is the human video doing the work, not the robot data.

[hosan @hosan300]

@_krishna_murthy Open tactile hardware is what data collection needs most right now. Every rig has its own sensor and none of the data lines up.

[hosan @hosan300]

@WaldenRobotics @RussTedrake Looking forward to this one. Curious if factory data ends up being the moat or the bottleneck.

[hosan @hosan300]

@sabina_beleuz Respect to them for showing it live. Donut-in-a-box is exactly the long-tail case demos skip and real data has to cover.

[hosan @hosan300]

@DexmateAI @SniperPaper @yi_wu_51 Only 20–36% of the time is the robot actually moving. Wild. Do faster models shrink that, or does the agent just think longer?

[hosan @hosan300]

@stoyanpstoyanov Currently manipulating data to help improve physical AI training

[hosan @hosan300]

Letting real-world failures decide what data to collect next is exactly right. Congrats on the launch.

[Paige ✈️ CoRL @paigeinsf]

Today we're launching @se3labs, backed by @ycombinator.

Robot models are getting better fast, but there's still a big gap between a great demo in sim or the lab and a robot that works reliably on a real task.

We're building real-world RL environments for robots. We start with tasks from industry partners, run policies on real robots, and find exactly where they fail. Those failures decide what data we collect and curate for the next policy update.

Next, we're bringing this to partner sites with physical sandboxes for real-world RL, so policies can train on the exact tasks and conditions they'll face in practice.

If you’re training robot policies or have a real task you want robots to do, we’d love to talk: https://t.co/uRUpPFWQKT or DM me directly!

See you at CoRL 🤖👋

[hosan @hosan300]

@yukez One model across grippers, hands and humanoids is the real flex. Would love to know how much of it came from human video.

[hosan @hosan300]

@TheHumanoidHub Teleop at a booth is still data. Bet they're logging every snack handoff.

[hosan @hosan300]

@PrismaXai @corl_conf Anything with deformables. Cables, bags, a shirt that's inside out. That's where the data gets expensive.

[hosan @hosan300]

@SereactAI @sigmundhh A robot retrying a slipped grasp on its own beats a perfect demo. That's exactly the stuff teleop data never shows.

[hosan @hosan300]

@chris_j_paxton Escaping the cage might be the best benchmark at CoRL this year.

[hosan @hosan300]

Every world model at CoRL this week wants the same thing: lots of human video of hands doing real work. Data collection just got a lot more interesting.

[hosan @hosan300]

1,000 hours of first-person human video beat 5,000 hours of robot data in messy scenes, per the new Open-HumanWAM paper. The cheap data is quietly becoming the good data.

[hosan @hosan300]

@Mr_TuanHoang Writing down what done looks like fixed most of my agent problems. The rest was me changing my mind halfway.

[hosan @hosan300]

@chris_j_paxton Waking up to a tidy house is honestly the best robot pitch there is. Way better than another backflip.