←
Ishank Gupta3,307 posts
Ishank Gupta avatar

Ishank Gupta

@ishank20

Co-founder @Humynlabs | Building the research & data engine for Physical AI

◷ Joined May 2009▧ 236 Photos and videos
Ishank Gupta avatar
Ishank Gupta@ishank20·4h

A robot’s ability to do a task is not the same as its readiness to be trusted with the job. That gap does not close in a lab. Today we announced $47M in Series A funding for Humyn Labs, led by @Accel, to close it: capture people doing real work in 20+ countries, turn it into training-ready signal, then put robots to work in those same environments and feed every shortfall back into the next model. Six months in: $50M+ annualised revenue, frontier labs and humanoid companies as customers. The next generation of robotics will not be decided by models alone. It will be decided by whoever gets those models trusted in the real world. That is the job. DMs open.

We're building the infrastructure that takes Physical AI out of the lab and into the real world. And we've raised $47M to make that happen. Physical AI models are getting better every month. But getting robots to work reliably in real-world environments, across full shifts and changing conditions, is a different challenge altogether. The missing piece isn't just better models. It's deployment readiness. At Humyn Labs, we're building the research and deployment layer to close that gap. Here's how: - Enrichment Engine: Our end-to-end data engine transforms real-world work into labelled, training-ready signals with verified provenance. - Deployment Readiness: We work with labs and site operators to deploy robots in those same environments, identify where they fall short, and feed those gaps back into the next generation of models. - Capture Hardware: Our own capture devices across sight, sound, motion and touch record how people actually perform a job, built to each lab's specifications. We've made significant progress since launch: - $50M+ in annualised revenue within six months. - Capture and enrichment operations across 20+ countries in Latin America, East Asia, South Asia and Southeast Asia. Today, we’re announcing $47M in Series A funding: - $26M in equity, led by @Accel with participation from @Prosus_Ventures , @Bajaj_Finserv Intelligence through Bajaj Finance, Bajaj Alts AI Opportunities Fund, BeeGlobal, @wearedallasvc, @insights_by_LC , @jumptrading and a marquee angel investor. - $21M in venture debt, led by @LighthouseCanton with participation from @TrifectaCapital. With this funding, we're expanding our deployment readiness infrastructure in Brazil, South Korea and India, scaling our capture hardware and enrichment engine, and growing our teams across San Francisco, London and Seoul. Our mission is simple: accelerate the deployment of Physical AI in the real world. The next generation of robotics won't be built on models alone. It'll be built on the infrastructure that gets those models working in the real world. Learn more: humynlabs.ai

1:17
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Nov 12

On benchmarks, hand tracking looks solved. On real work, it breaks. We tested 4 open models (WiLoR, HaWoR, EgoForce, Dyn-HaMR) on people doing real jobs. Here's what broke, why, and what we built to fix it ↓

Post attachment
Pushpak Agrawal avatar
⟳ Ishank Gupta retweeted
Pushpak Agrawal@pushpak_agr·Nov 6

wanted to test this intuition from a longtime, finally some output. used egodex dataset and retargetted to @i2rt_robotics yam bots. some amazing work at @humynlabs

0:07
0:08
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Nov 6

A few moments from Robotics, Served 🍽️ Last night, we brought together some of the top minds working in physical AI for dinner in Bengaluru, hosted by The Embodied AI Collective, Humyn Labs’ researcher community. We talked about everything from robotization and deep sciences to stories of working in different parts of the world. Mostly though, it was people finding genuine interest in each other and having a good time. Thank you to everyone who took the time to join us. We’re glad we could share a table. @ishank20 @manishdiesel @MaiDiab_

Post attachmentPost attachment
Ishank Gupta avatar
Ishank Gupta@ishank20·Nov 3

Not surprised. And Physical AI has no internet to scrape.

Interesting to see a big jump in data vendor spend this past quarter for Anthropic. Much more in my Anthropic S-1 breakdown which goes out shortly.

Post attachment
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Nov 2

Teaching robots is tougher than it looks. 👀 We asked Physical AI researchers and tinkerers to explain their work. And ofc, what annoys them. Here's what they had to say ft. Devarshi Aggarwal, Shivam Kumar Singh, Ritul Singh, Pranshu Malik, Sahana Ameen, Pranav Prashant Shewale, and Lekhansh Rana.

0:56
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 31

Not every second of egocentric factory footage is useful training data. The hard part is telling which seconds are, at scale and with reasons. Humyn-WISE (Work Interval and Session Evaluator), developed by our in-house research team, labels each second of egocentric video as active or idle relative to the task in progress. It also explains each label: operating the fixture, fetching material, waiting on a machine, or watching a coworker. How we evaluated it We scored Humyn-WISE and six frontier multimodal models second by second against a human reviewer on a 10-minute factory clip. Every model received the same prompt and the same phase table. Idle is the positive class, and seconds where the wearer’s hands are out of frame count as idle. Results (F1 · MCC · Cohen’s Kappa · IoU): Humyn-WISE: 85.4 · 79.0 · 78.9 · 74.5 Gemini 4 Argon: 79.7 · 74.0 · 72.7 · 66.3 GPT-6.1 Sol: 77.7 · 67.8 · 67.7 · 63.6 GPT-6 Astra: 76.8 · 67.9 · 67.7 · 62.4 Gemini 3.8 Flash: 71.3 · 58.8 · 54.1 · 55.4 Claude Fable 5.1: 67.4 · 54.1 · 54.0 · 50.9 Claude Opus 5.5: 30.7 · −4.5 · −4.4 · 18.2 All scores in %. Humyn-WISE led on every metric. Why it matters At millions of hours, a few points of accuracy become thousands of hours of footage wrongly kept or wrongly cut. That gap decides whether a dataset is ready for robot training as delivered or needs cleaning first.

0:55
Ishank Gupta avatar
Ishank Gupta@ishank20·Oct 29

BTS from the lab. Calibrating our stereo cams by hand was landing around 0.7 px reprojection error. Good is under 0.5. So the team put a YAM arm on it. It moves the camera through a fixed set of poses in front of a ChArUco board, nobody touching it. Same poses every run, so the capture set is well-conditioned instead of depending on who held the camera that day. ~0.3 px now. Two minutes a rig. Nobody asks about calibration. Everything downstream inherits its error.

0:10
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 28

Meet the validation engine behind every Humyn Labs dataset. Built by our Director of Data, @MaiDiab_ , and her research team, it runs every hour of egocentric footage through six stages, from calibration to final file check, before it’s allowed anywhere near a model. On a 1,000-hour sample, 647.8 hours came out as verified training data. 352.2 hours were stopped, each quarantined with a logged reason. The surprise: half of what we rejected wasn’t broken. 177.9 hours were technically clean but had no manipulation signal, with hands out of view, idle, or too short to use. File-level checks would have passed all of it. In physical AI, data that only looks right teaches the wrong motion. So the bar is simple: if we can’t show an hour is right, it doesn’t ship.

Post attachment
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 27

This is what Bengaluru’s physical AI scene looks like in one room. Last Saturday, 30+ researchers and engineers joined us to work through one question: why does egocentric hand tracking break in the wild? They challenged each other’s assumptions, disagreed openly and took the problem further together than any one of them could have alone. We’re building more rooms like this, where top minds in physical AI share their work, debate open problems and build together. @ishank20 @manishdiesel

0:24
Aachman Maheshwari avatar
⟳ Ishank Gupta retweeted
Aachman Maheshwari@aachmannnn·Oct 27

first one in the books. thank you @ishank20 @manishdiesel for trusting us with it. the next room is going to be even crazier 🔥

This is what Bengaluru’s physical AI scene looks like in one room. Last Saturday, 30+ researchers and engineers joined us to work through one question: why does egocentric hand tracking break in the wild? They challenged each other’s assumptions, disagreed openly and took the problem further together than any one of them could have alone. We’re building more rooms like this, where top minds in physical AI share their work, debate open problems and build together. @ishank20 @manishdiesel

0:24
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 26

30+ researchers, 7 open questions on egocentric hand data. Here’s one of them: If hand reconstruction were perfect, what would still stop human video from training robot hands? We're writing up where the room landed. What would you pick, and why?

Post attachment
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 25

Yesterday we hosted our first closed-door research session in Bengaluru. 30+ researchers from ARTPARK (IISc), Samsung R&D, TCS Research, Siemens, Lossfunk and robotics startups joined us for the evening. One question: Why does egocentric hand tracking break in the wild? We started from HaWoR (CVPR 2025), looked at real failure cases from our data, and worked through open questions together. Lots of discussion, with real takeaways. We're writing up what the room agreed on into a report, and will share it soon. The idea is simple: get the right people in one room, sharing what's working and what isn't. More sessions, more cities, more events coming. Thanks to everyone who came, and to Dr. Sandeep for leading the room :)) @manishdiesel @ishank20 @aachmannnn @omjambhale3

Post attachmentPost attachment
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 21

This Saturday, our head researcher Dr. Sandeep Reddy Sabbella is leading a closed-door session at our Bangalore office on why egocentric hand tracking breaks. He's seen this problem from both ends. During his PhD at Sapienza, field data was extremely difficult to collect with environmental changes and agricultural crop cycles. Hence they built VR simulation to replicate the data that was collected for conducting experiments. At Humyn Labs, he works with first-person footage of hands cooking, folding and fixing things. Hands leave the frame, wrap around objects, and the camera never sits still. He'll share how we're tackling it. Then India's top researchers in hand pose, egocentric vision and robot learning take it from there. Sat, 24th Oct. Indiranagar. Last 5 spots remaining. Grab yours below 👇

Post attachment
Humyn Labs avatar
⟳ Ishank Gupta retweeted
Humyn Labs@humynlabs·Oct 20

Physical AI won't work in a lab. It'll work in the real world. So we build its training data from real people in 20+ countries, home to 39% of the world's population. Hiring engineers and researchers to scale it to petabytes of vision, speech, touch and motion data. Roles ↓

0:08
Ishank Gupta avatar
Ishank Gupta@ishank20·Oct 20

Saturday. 30 seats. Apply now.

We’re hosting a closed-door research session in Bengaluru this Saturday, with researchers from India’s top labs working on hand pose, egocentric vision and robot learning. The core question: Why does egocentric hand tracking break in the wild? Over the evening, we’ll: - Map where hand tracking fails using real cases - Check whether today’s benchmarks measure what robots actually need - Share how we’re approaching it, and learn together The session will be moderated and led by our Lead Researcher Sandeep Reddy Sabbella, PhD. What the room figures out shapes our next round of egocentric data. Sat, 24th Oct. Indiranagar. Limited to 30 researchers. Apply below 👇

Post attachment