Skild AI says it has surpassed $100 million in annual recurring revenue just 10 months after beginning commercial deployments of its robotics AI technology.
The company now has more than 60 paying customers using its technology for applications including material movement, deliveries, site inspection, security, food preparation, warehouses, factories and data centers.
Deepak Pathak, co-founder and CEO of Skild AI, and Abhinav Gupta, co-founder and president, say the company has also recognized $50 million in revenue since deployments began 10 months ago.
Around 90 percent of Skild AI’s revenue comes from robotic manipulation applications, according to the founders. Mobility accounts for approximately 10 percent, including autonomous mobile robot solutions associated with Zebra Technologies and Fetch Robotics, which account for 4 percent.
The figures were disclosed by Pathak and Gupta in a blog post outlining the company’s emphasis on real-world deployment rather than robotics demonstrations.
“From the beginning, we have focused on deployment as a core part of the technology itself and not the outcome of it,” they write.
Nvidia, Foxconn and Sumitomo deployments
Skild AI is working with Nvidia and Foxconn to deploy its Skild Brain technology on dual-arm robotic manipulators for high-precision assembly of Nvidia Blackwell systems.
The company says these manufacturing processes change with each product cycle and have traditionally required robots to be reprogrammed as production requirements change.
Skild is also working towards deploying its S1 robotics model at Sumitomo Wiring Systems to automate processes in wire harness manufacturing that have previously proved difficult to automate.
In another project, the company is working with Mitsui & Co to pilot S1-powered general-purpose robots in commercial kitchens. Mitsui’s associated supply chains serve 1.4 million meals per day across Japan, according to Skild.
Pathak and Gupta argue that these deployments are important not only commercially but also as a source of information for the company’s robotics research.
“Deployment is the hidden pillar of robotics research because it’s where robotics happens. There is no substitute,” they write.
‘Seeing is not believing in robotics’
The founders also question the robotics industry’s reliance on demonstration videos as a measure of technological progress.
They point out that a successful video clip can look similar regardless of whether the underlying robotic system achieves the task reliably or succeeds only occasionally.
“We taught our model to make eggs last year. It took a week to cook the first egg, then two months to make it work reliably with different eggs and in different setups,” Pathak and Gupta write.
“This is why seeing is not believing in robotics. Deployment is what matters. The amount of effort to squeeze the last 5 percent of performance greatly exceeds that of the first 95 percent.”
The same distinction applies to the speed at which robots complete tasks, they say.
In a production line containing both human and robotic workstations, the slowest station can constrain the throughput of the entire operation.
“A robot that is 99.9% accurate but ten times too slow is not almost deployable. It’s not deployable,” the founders write.
“A deployment-first company optimizes for success under time constraints from the beginning.”
Learning from changing production environments
Another problem encountered during deployment is the frequency with which manufacturing processes and working environments change.
Suppliers can change components, factories can rearrange workstations and customers can alter assembly processes after a robotic system has been installed.
Skild says continually collecting new datasets and retraining models whenever such changes occur would make widespread deployment difficult to scale.
The experience influenced the development of S1, the company’s latest robotics model.
S1 is designed to learn a new task from a single video example using in-context learning. An operator can provide the robot with a new demonstration as a prompt without updating the model’s weights.
The model is built using Nvidia AI infrastructure to support training across Skild’s robotics datasets.
“Would this have become such a priority if we had stayed in research land? We don’t think so. Deployment taught us to ask this of ourselves, and S1 is our answer,” Pathak and Gupta write.
Moving beyond robotics demos
The founders also argue that robotics companies can find it difficult to maintain both a culture focused on demonstrations and one focused on production deployments.
“You can cherry-pick a good demo from a robot with 50% accuracy. To deploy, you have to work on the other 50 percent – which is much, much harder,” they write.
“If demos are what a company rewards, demos are what people will work on. We chose to reward deployment.”
Skild ultimately intends to use information generated across individual deployments to improve its general-purpose robotics models, creating what it describes as a “deployment data flywheel”.
The approach begins with a foundation model trained across different robots, tasks and embodiments. The model is then deployed into individual applications and adapted to specific tasks.
Data and experience from these specialized deployments can subsequently be incorporated into the broader model, potentially reducing the amount of specialization required when it is deployed in new applications.
“The next deployment starts from a stronger model,” Pathak and Gupta write.
“That’s our strategy. By deploying, we’re making the ‘high-school self’ – S1 – better and better. The more the generalist learns, the less specialization the next deployment should need. This is the deployment data flywheel.”
The founders conclude: “The era of demos is over; the era of deployment has begun.”