Pranay Gupta

Chief Robot Officer @ | Prev: ML @ Zomato, Quant @ Da Vinci, ML @ MasterCard | CSE, IIT Delhi

[Pranay Gupta @PranayGupt97167]

We trained a robot policy to run at ~4x faster than the teleop data they were trained on. It held >95% success rate across 500 real-robot runs. Here’s a thread on our first blog at https://t.co/OE1dgDUduB 🧵

[Pranay Gupta @PranayGupt97167]

Robotics will eventually be solved by robots deployed in the wild, learning from themselves and the humans around them.

[Deepanshu Rohilla✈️CoRL’26 @Deepans36819800]

Something unreal happened yesterday.

Our robot learned how to close a dishwasher by itself.

The robot learned this purely from deployment data. The robot had never seen this particular task before. Every time, a human had to step in and finish the task.

Over time, with the deployment experience, the robot learned to do it on its own. Now, as it does more of this, it will speed itself up. @PranayGupt97167 wrote a nice article about this on X

We are deploying at scale across kitchens. If that sounds interesting, reach out to us!

[Pranay Gupta @PranayGupt97167]

Did not expect to wake up to @Google copying our name

[Entrepreneurs First @join_ef]

Yeok landed in San Francisco and, two days later, flew to Taiwan to get Caldera’s motors built.

He slept in bunk beds with factory workers, took cold showers, and stayed focused on getting the job done.

He and Raghav met through EF. Now they’re building Caldera with one goal: to build the best robotic manufacturing company in the world.

Next week, they’ll be pitching at Demo Day in San Francisco.

[Sanyam Jain @SanyamJain36928]

Ordered veg pasta, got chicken ate some.
On the auspicious day of Paryushan Parva !

I really trusted swish but this is such a huge let down.

This has hurt my religious beliefs ! Is this the kind of service you want to provide @ujjwal_sukheja @aniketshah30 @fssaiindia !

[Pranay Gupta @PranayGupt97167]

@QuanVng We only have access to pi0.5 and ABC at the moment.

In talks with your team to get access!

[Pranay Gupta @PranayGupt97167]

We trained a robot policy to run at ~4x faster than the teleop data they were trained on. It held >95% success rate across 500 real-robot runs. Here’s a thread on our first blog at https://t.co/OE1dgDUduB 🧵

[Pranay Gupta @PranayGupt97167]

5/5

All of this combined, our MTT (median task time) for bimanual handover dropped from 37.3sec to 9.4sec at >95% success rate running π0.5.

Full writeup: https://t.co/w8BHULAMKh

[Pranay Gupta @PranayGupt97167]

4/5

Lever 3: DAgger intervention data adds accuracy but we have seen that Sped-up DAgger intervention data added another 10-15% to speed as well.

[Pranay Gupta @PranayGupt97167]

Just need 10 more claude sessions to actually be productive

[Deepanshu Rohilla✈️CoRL’26 @Deepans36819800]

Not jealous of @PranayGupt97167 ‘s setup

[Pranay Gupta @PranayGupt97167]

@r_hul_parmar @getRilo @Adobe Congratulations to @Adobe on getting an amazing set of people

[Deepanshu Rohilla✈️CoRL’26 @Deepans36819800]

Just make it exist first, you can make it good later

Crazy how customers were already loving the May version

[Pranay Gupta @PranayGupt97167]

@awsdevelopers I’m not but aws probably is

[Pranay Gupta @PranayGupt97167]

Google having the worst ai model is genuinely insane considering they had the most data, compute, resources and they invented the transformer architecture

[Pranay Gupta @PranayGupt97167]

@0xIlyy

[Pranay Gupta @PranayGupt97167]

Really wish I was studying chip design instead of playing farmville when I was 10

[Pranay Gupta @PranayGupt97167]

@CocaCola we will run our arms in the @DietCoke plant for free, I cannot survive another shortage😭