For over a decade we’ve been deploying robots into the wild. We give physical AI the verified real-world data—graded at capture—to take models beyond demos.
Joined October 2018
- Tweets553
- Following1K
- Followers994
- Likes369
Most training data ships as a clip grid and a promise.
Formant Vetted grades every take at capture, against your spec.
464 takes: 187 meet spec, 277 don't, and you see exactly where the gaps are.
Know what's usable before you pay.
Quality you can prove. Diversity you can measure.
formant.ai/real-world-data-f…
#PhysicalAI #RobotLearning
Why would a food delivery robot need a $10,000 robotic arm? 🤖👇
Hint: It's not to hand you your pizza.
@trossenrobotics CEO Anish Shah on why "simple" tasks like pushing crosswalk buttons are where Physical AI is deploying today.
📺 Full episode: youtu.be/xn9E7sRiCMk
From Stanford ALOHA to production fleets: how Trossen is solving the robotics data problem. #PhysicalAI #Robotics
ALOHA made robot-learning hardware affordable. Now everyone can collect physical AI data, but nobody knows what good data looks like.
Our host @henrikbennetsen sits down with Anish Shah, CEO of @TrossenRobotics 👇
What’s your #1 data-quality red flag? 🤔 #TrainingData
Formant retweeted
A mature approach to autonomous systems balances technical performance with ethical considerations under resource constraints.
Formant retweeted
Robotics Data Collection companies are the Waymo safety drivers of today.
They serve an important but very temporary role.
Some parts of my conversation with @FormantInc on data
Formant retweeted
prominent companies doing this:
> Formant (@FormantInc)
> InOrbit (@InOrbitAI)
lots of untapped opportunity still remaining in this field btw
As robotics manufacturers seek to make their software + hardware increasingly proprietary, operators will find it increasingly difficult to manage fleets of different robots that aren’t all the same brand
As we reach embodied AGI, it is critically important that we ensure interoperability + common standards for everyone.
10–20 raw hours to get 1 usable hour of robot data. Quality is the new bottleneck. Notes from NVIDIA’s Data Masterclass 👇 #LongHorizon #PhysicalAI
A year ago, the ask was “simple”: head-mounted video of everyday household tasks.
Pilot: a few hundred hours.
Today: 60,000 hours a month.
@kmgujarati of @awign_india on how fast #PhysicalAI data is scaling.
Like what you see? Check out the full episode here: youtu.be/iLosm1msJx4?si=EpCa…
#PhysicalAI #RoboticsTrainingData
Quiet on the outside. Building on the inside.
Introducing Formant Vetted: audit before you buy.
Quality-checked training data for physical AI.
Stuck on #RoboticsTraining data? We can help ⬇️ formant.ai/real-world-data-f…
"Another day, another Physical AI conference at the beautiful Fort Mason" — @henrikbennetsen , CBO.
We’re at @encord_team's #LongHorizon today! Want to discuss training data strategies and scaling physical AI?
Grab time with Henrik & Anne here: calendly.com/anne-formant/30…
Is the open web tapped out for AI training?In our latest episode with @SignalFire investor @RyanWexler,
we break down:
🔹 Shift from scale to hyper-specificity in post-training data
🔹 Phase 2 data strategies: selling to frontier labs vs. keeping data internal
🔹 The economics of robotics training data & teleop
🔹 Why physical AI needs its own evaluation frameworks
#StartupStrategy #AIFounders #PostTraining #DataMarket #DeepTech #SignalFire
🎧 Catch the full episode: youtu.be/RoRJNv90iMQ?si=HASf…
"A robot policy that fails 20% of the time is unusable on a factory floor."
On the latest episode, We chatted with Miki (@mikisafr), Co-founder of @NorthstarRobotics, about what it actually takes to bridge the 70% ➔ 99.999% reliability gap in Physical AI.
🧵 Here are 5 takeaways on failure recovery data, DAGGER interventions, and why lab demos don't survive production:
#PhysicalAI #Robotech #RoboticsTrainingData
Success data teaches robots what to do. Failure data teaches them what not to do. In physical AI, failure data isn’t a bug—it’s gold. 🪙
Great insights from @mikisafr at @NorthstarRobots and why edge cases and real-world mistakes are the most valuable training data in robotics.
What are your thoughts on what makes training data move the needle? Drop in the comments below 👇
Stay tuned for the full episode coming out tomorrow.
#RoboticsTrainingData #PhysicalAI #Formant
#PhysicalAI is moving faster than anyone predicted.
In just 3 months, we’ve seen:
Massive pre-training data buyouts
Expanding lab budgets & contract sizes
The rise of "neo-integrators" pairing off-the-shelf hardware with open-source models
Is deployment finally here? 🤖🔥
Check out what @mikisafr from @NorthstarRobots had to say on what's driving it. Stay tuned for more insights.
Formant retweeted
What goes into one hour of useful data when there is hardware involved?
A piece of my conversation with @henrikbennetsen from @FormantInc
When @theSamPadilla shut down EIDON AI, a hardware and data startup building training data for physical AI, he and his co-founders open-sourced the company's data, hardware designs and firmware. Then Padilla published a post-mortem called Robotics Data Is a Broken Business that had a viral moment last week.
We invited Sam to share his hard earned wisdom on our podcast: youtube.com/watch?v=cZAd8y5y…
Check out Sam's piece here nitter.cf/theSamPadilla/status/2…
After 2+ years in the robotics data space, we are shutting @Eidon_AI down.
The thesis was right. But the business is brutally hard.
We close this chapter by open-sourcing everything we built and sharing lessons for anyone venturing into the space.
@theSamPadilla continues to share important truth in his follow up piece from today.
nitter.cf/henrikbennetsen/status…
The section on Specificity gets at how we need training data to show how skilled people actually do economically important work. Check out @theSamPadilla interesting follow up piece
nitter.cf/theSamPadilla/status/2…