The robotics companies that win won’t just have the best demos. They’ll have the strongest deployment loop. Skild AI: first commercial deployment → 60+ paying customers → $100M ARR in 10 months. Every deployment creates new problems to solve and new knowledge to feed back into the system.

Have seen a few people (myself included) try to present data needs for physical AI as a pyramid, but this framework from @deepakpathak does a much better job at showing the pros and cons of each Paraphrasing Deepak: “You cannot rely on one data source to scale contribution… A combination is what will solve the data robotics problem” Congrats on the $100M ARR announcement @SkildAI !

For real world robotics deployments, learning the initial procedure is only part of the challenge. Most startups' robot foundation models are already able to do that. To deploy in to the real world, robot also needs to adapt when that procedure changes. In-context learning offers a way to communicate those changes through another demonstration, rather than automatically starting a new data-collection and training project. That reduces the engineering required to keep complex deployments useful as the customer’s operation evolves. Skild had a unique focus on in-context learning, and it is reaping the rewards in the scale of their deployments within the first year.

Great article from the Co- Founder of @SkildAI "In robotics, you cannot leave deployment until the end. How will the supply chain work? Who installs it? Who owns the integration? Who fixes it when it breaks? Who updates it when the process changes? You don’t fully understand these questions until you’re there. Deployment is the hidden pillar of robotics research because it’s where robotics happens. There is no substitute." I think its interesting to see Skild is using similar logic and deployment method as $CCXI / @agilityrobotics. To a certain degree, it feels like Agility "forced" Digit into deployment, focusing on commercial traction rather than perfecting all of the technological components. With that said, they have the highest level of deployment, the value of which we may not fully understand until humanoid manufacturing scales. The "Data Flywheel" that Agility hypes is something I've touched on in my piece here:

How did a $100m ARR robotics company solve scale? "Unlike language models, speech or video, in robotics there are three axis to evaluate data quality. One is how scalable you are, how diverse it is, and how close it is to robot." Robotics has four ways to collect training data. Each one comes with its own fault. Robots collecting their own data produces the best-targeted data and scales slowest, because the physical world will not run faster than real time. Teleoperation is the industry default and produces almost no diversity. Simulation runs far faster than real time, but every new task needs an environment built by hand. Human video is the most abundant and the furthest from a robot, since a person has a different body and you cannot see the forces they apply. There is not one "Golden path" as @deepakpathak put it on stage at AUTONOMOUS earlier this year. @SkildAI

A banger memoir by @deepakpathak. @SkildAI had successfully been able to build something that was once a distant dream for robotics engineers. The important part is they got it right after doing lot of things wrong! That’s how it must be.

“If every change requires collecting a new dataset and running another round of post-training, you’re signing up to repeat that work for as long as the robot is deployed. This is not scalable.” Adaptation Cost has always been the right metric for generalization. You must minimize it through diversity.

Robots need to be useful and need to be out in the world doing a variety of stuff to get better. They can't just learn in labs. In the end the two things that stick out are - a robot is a tool. It has to actually do a job, and if youre not trying to use it for that job it probably doesnt work - along the same lines, what you see is what you get -- when you see a demo you have to remember its the absolute best the robot has ever done, and if you didnt see it do something, you must assume it cant do that thing

@SkildAI hit $100M ARR about 10 months after first commercial deployment, with 60+ paying customers and hundreds of robots in the field. @deepakpathak’s line is the right one: demos are easy to film, deployments are the research.

Skild, a fast-growing startup that develops software to help robots learn tasks, has reached $100 million in recurring revenue run-rate, 10 months after beginning commercial use

A well articulated piece on how @SkildAI approaches deployment: 1) rewarding deployment rather than demo to build the right company culture 2) think of S1 as the high school brain and each deployment helps it build the specialized knowledge Congrats @deepakpathak and S1 team on this milestone!

The best thing you’ll read all day. Something @deepakpathak and @gupta_abhinav_ have always stressed at Skild is that not only is research a necessity to develop the best deployment solutions, but deployment is a necessity to contextualize, evaluate, and guide research developments. In robotics, the two must work hand in hand to be most effective. The whole reason for moving from academic research to industry is to accelerate that feedback loop between research and deployment. A method developed in the lab isn’t good enough until it’s proven in the real world. That’s the whole point of research, to develop solutions to problems that actually matter, and unlike digital agents, capabilities of physical agents cannot easily be verified in isolation. We believe progress towards making robots a viable solution means getting your hands dirty with real deployments in real world environments for real customers. Doing so often involves a fair share of grunt work and a different level of discipline that turns off many researchers. It’s easy to feel like the work is done and it’s time to move onto the next cool dexterous capability once you’re able to shoot a video of the result. It’s an entirely different mentality when 1 failure out of 1000+ trials is enough to keep you up at night.