Varun Agarwal

Building in #PhysicalAI #Robotics

[Varun Agarwal @varunagarwal971]

@manumerous @enact_intel congrats Manuel and team 🎉 sensing force through the actuators instead of bolting sensors onto fingertips is such a clean bet. really excited to watch this one

[Varun Agarwal @varunagarwal971]

Looking for a multimodal data annotation partner in India. To be used for robotics use cases. Pls reach out
#dataannotation #multimodal #robotics

[Varun Agarwal @varunagarwal971]

@cmarschner fair q, review is the real bottleneck right now. but imo the diff is becoming the cosmetic check: a powder coated part can look perfect and still peel later bc the pretreatment was off. do you review the agent's plan too, or just the code?

[Varun Agarwal @varunagarwal971]

@lukas_m_ziegler @Siemens @nvidia let's go, mad week. one thing I'd watch: an electronics line is the friendly case, high volume and tight fixtures. the long tail is high-mix shops where every batch brings part variation and fixtures get rebuilt. is the testbed tackling that, or mostly repeat runs?

[Varun Agarwal @varunagarwal971]

@GlenBerseth building all four on the same data is the right way to settle it. my bet: for contact-rich tasks none of them win on pixels alone. in powder coating the gun path looks fine on camera and coverage still fails in the recesses. does the data carry force or outcome labels?

[Varun Agarwal @varunagarwal971]

@chooi_jeq mad pace. the catch imo: most tasks score whether the motion finished. in spray painting a pass can look done and still fail QC on coating thickness or runs at the edges. any plans for a tier where the score is the downstream quality check, not task completion?

[Varun Agarwal @varunagarwal971]

@mangahomanga love this framing. imo feasibility is the right lens, but in contact-rich work the intent lives in force, not motion. sanding a curved panel, the path looks the same at the right pressure or when you burn through. can the constraints encode contact force, or only kinematics?

[Varun Agarwal @varunagarwal971]

@jang_yoel @scott_e_reed @youliangtan @GeYunhao congrats on the launch. agree generalization wins, but imo the 90 to 99.9% part is where factories live. in weld prep the same fixture sees parts off by a few mm every batch and a miss is rework. can Dream-1 absorb that variation from a few demos, or is it retraining?

[Varun Agarwal @varunagarwal971]

@YaseenTech4 let's talk!

[Varun Agarwal @varunagarwal971]

@IlirAliu_ sanding is the sneaky hard one. a good finisher changes pressure and angle by feel as the surface changes, and none of that shows up on camera. would love to see it on edges and patch repairs, not just flat wall. that's where the multimodal data gap shows up

[Varun Agarwal @varunagarwal971]

@arvindakaraja @michellelsun @arvindakaraja we should talk

[Varun Agarwal @varunagarwal971]

@antopatrex1 8% drift after 4 hrs is crazy and nobody talks about it. 20-min demos vs a full shift are basically different datasets, same for human demos. fatigue changes technique

[Varun Agarwal @varunagarwal971]

@KGeN_IO this is so real. same motion, different station, totally different task. we see it constantly with skilled hand work. capturing that variety is the whole game imo

[Varun Agarwal @varunagarwal971]

@omtcyang riding on production instead of disrupting it is the move. real q: how far does one ego cam get you on contact-heavy steps before you need more sensing? we keep hitting that on shop-floor tasks

[Varun Agarwal @varunagarwal971]

@joshavata @tkkong @philhchen 94 vs 71.5 on actions is wild. captions are easy, nailing the contact + force moments is where it gets real. hitting the exact same wall capturing skilled hand work right now. would love to nerd out on this

[Arvind Jain @jainarvind]

In many discussions, I find that context is often being referred to as a synonym for more data, but more data doesn't necessarily mean more context.

Organizational theory is a good way to think about context, which breaks enterprise knowledge into four categories: know-what, know-how, know-why, and know-who.

Take a sales deal.
𝗞𝗻𝗼𝘄-𝘄𝗵𝗮𝘁 is the account plan, pricing, security questionnaires, and redlines.
𝗞𝗻𝗼𝘄-𝗵𝗼𝘄 is how to move the deal forward, like when to involve security and how to sequence the work. 
𝗞𝗻𝗼𝘄-𝘄𝗵𝘆 is rationale, like which objections indicate real risk and which approvals are routine. 
𝗞𝗻𝗼𝘄-𝘄𝗵𝗼 is the social knowledge, like who decides, who has handled a similar issue in the past, and who needs to be involved.

Most of what AI can retrieve today falls in the know-what category. The rest is trapped in scattered conversations, unwritten routines, and informal human networks.

We've spent the last 7 years at @Glean building toward this gap. Our context layer to connect know-what, know-how, and know-who, while inferring know-why from how work actually happens.

An AI tool with access only to know-what can find the account plan but will fall short in telling you who to involve, which objection deserves attention, or what to do next.

It’s the difference between knowing the account, and knowing how to move it forward.

[Varun Agarwal @varunagarwal971]

Why is there not enough talk about the missing context layer in Agentic systems?
AI agents need way more context than they currently have.
Simply having a markdown file to store context does not cut it.

[Varun Agarwal @varunagarwal971]

@jonstephens85 @LightwheelAI Will be there too, would love to chat

[Varun Agarwal @varunagarwal971]

@danfei_xu Will be there too

[Varun Agarwal @varunagarwal971]

@manju_kumar_ Will be there too, let’s connect