Pushkal Katara

@pushkalkatara

SkildAI | CMU

San Francisco, CA

pushkalkatara.github.io

Joined February 2017

1,149 Following

364 Followers

Replies Reposts Media

Pinned
Pushkal Katara
@pushkalkatara
Aug 25
What an incredible time to join @SkildAI Every robot deployment so far has been gated on collecting data for the exact task you want. Hours of teleop, fine-tuning, specialist policy that does one thing. That's why robots look great in demos and stall in the real world. S1
Skild AI
@SkildAI
Aug 25
Introducing S1, our new foundation model that learns from one example. It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning. Watch S1 operate in real-time via in-context learning:
00:00
4
Pushkal Katara
@pushkalkatara
Sep 23
Surprising how well self-play works for post-training robot policies. ~140 years of physical experience, all from playing against itself in sim. That scale of experience is hard to collect through teleop (too expensive to scale) or video (no physical priors). Soccer is just the
Skild AI
@SkildAI
Sep 23
We trained a robot to play football. How? Self-play for 140 years in a virtual World Cup. Meet the #Messinator:
00:00
Pushkal Katara
@pushkalkatara
Sep 10
building data flywheel is one of the most important bit towards general purpose robotics!
Skild AI
@SkildAI
Sep 10
We just hit 100M ARR within 10 months of starting deployments. We are in factory lines. On construction sites. In kitchens. In data centers. Cleaning. Welding. Building. Cooking. Deploying.
00:00
Pushkal Katara
@pushkalkatara
Aug 25
Early evidence that in-context learning scales in manipulation. Show S1 one video and it performs tasks it was never trained on, up to ten minutes long. All from one set of weights, no post-training. Robotics is leaving the BERT era!
Skild AI
@SkildAI
Aug 25
Introducing S1, our new foundation model that learns from one example. It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning. Watch S1 operate in real-time via in-context learning:
00:00
Pushkal Katara
@pushkalkatara
Aug 15
Two incredible years at @genrobotics_ai have come to an end. Being here from the early days was the ride of a lifetime: hard problems, robots going from demos to real systems. Learned so much from @jonathanhuang11 @saihv, Dinesh, and @akapoor_av8r. Excited for what's next. 🤖