Council Post: In Robotics, Who Pays For A Robot's Education?

Laila Burns is CRO and Co-Founder of Sielo Robotics.

getty

​​Boston Dynamics spent months preparing two Atlas robots for a 90-second parkour routine. Their most challenging obstacle proved to be a vault over a waist-high beam; they succeeded at only half of the filmed attempts. During one attempt, a hydraulic hose burst from Atlas’ knee while it walked down three stairs, spraying pink fluid across the floor as the robot collapsed.

That was one advanced robot performing one rehearsed routine inside a controlled environment. A useful commercial robot must complete thousands of less predictable tasks without an engineering team on standby.

I’d like to share a framework I developed with our company’s CEO that examines four ways companies can pay for a robot’s education. Drawing on an analogy to human learning, he describes these models as “school,” “laboratory practice,” “early graduation” and “apprenticeship.”

School: The Company Pays People To Teach

The robotics industry often uses the school model and treats robot education like a university degree. In the school model, the company pays people to create lessons for the robot. Companies raise large sums and train machines in controlled settings, hoping they graduate to useful work.

Foundation models can make this teaching more effective. Physical Intelligence trained its model using internet-scale knowledge and data from seven types of robots. It can fold clothes and make coffee, although difficult tasks still require high-quality data from the specific robot and job.

Sunday Robotics built wearable gloves that remote workers wear while completing household tasks. The company uses those movements to train its foundation model. The company raised $165 million in March 2026 and expanded its data team before planned home deployments.

School gives teams control over what a robot learns, but every new room, object, mistake and recovery creates another lesson someone must produce and pay for.

Laboratory Practice: The Company Pays The Robot To Learn

In the laboratory model, the company pays for simulation, testing, repairs and operations while the robot practices.

Waymo is the strongest example. It began as Google’s self-driving car project in 2009 and received years of support before becoming a separate Alphabet company. In 2025, it completed 15 million rides and reached more than 400,000 paid rides each week.

This approach can work when the market is enormous and investors can wait for years. Most companies building farm robots, construction machines, assistive arms and industrial systems cannot fund every lesson before customers begin paying.

Simulation remains essential for studying dangerous situations and rare failures. However, it cannot reproduce every condition found in real use. Laboratories can control lighting, layouts, objects and network connections. Customers bring crowded rooms, reflective surfaces, worn parts, poor connections and goals engineers never anticipated.

Sergey Levine, a founder of Physical Intelligence and professor at Berkeley, argues that collecting real-world robot data is primarily an industrial challenge. Once robots are widely deployed, the challenge may shift from finding data to deciding which data matters.

Early Graduation: Deploy First, Improve Later

Some robots can begin independent work early because their first job tolerates mistakes.

The original Roomba sold for $199.95 and used simple navigation. It did not need a detailed understanding of chairs, rooms or household activity because vacuuming allowed many acceptable paths. Repeated movement still removed dirt, while most collisions caused little damage. By May 2006, iRobot had sold more than 2 million Roombas. The product generated revenue, customer feedback and real operating experience while the company continued improving it.

Roomba was able to graduate early because its first occupation was forgiving. Robots handling medicine, surgical tools, factory parts or someone’s dinner face far greater consequences when they make mistakes and need another path into real-world use before achieving full independence.

Apprenticeship: Customers Pay For Useful Work

The apprenticeship involves paid deployment in which a human still supplies part of the judgment or control. It must create value for everyone involved: The customer receives useful work, the company supports continued development, and the robot gains experience that improves performance. Otherwise, the pilot remains a research project or transfers the company’s burden to the user.

Intuitive Surgical built da Vinci around this model where surgeons provide medical judgment and control the tools, while the robot offers steady movement, improved vision and precision. Da Vinci systems have been used in more than 20 million procedures according to the company, meaning decades of experience inside hospitals and surgical workflows. This does not guarantee autonomous surgery, but it positions the company well because its machines already operate where future surgical intelligence must work.

Tesla followed a related path with customers buying useful cars as the company deployed cameras, computers, software, charging stations and service centers at scale. Waymo offers the stronger fully driverless service, while Tesla’s Full Self-Driving still requires an attentive driver. Tesla’s advantage is economic: Customers fund and use the cars while the company develops its driving technology.

Wheelchair-mounted robotic arms can follow the same model. Systems such as Kinova’s Jaco can be controlled through a joystick, head controls or sip-and-puff systems. Newer platforms, including my company’s robotic device, can solve immediate reaching and grasping problems before reliable household autonomy arrives. Real use reveals goals, corrections, failures and recovery methods that controlled demonstrations may miss.

This learning requires permission and safeguards, keeping failures safe and recoverable, collecting useful data deliberately and studying dangerous edge cases purely through simulation.

Understanding Each Pathway

Strong robotics companies typically will use all four models. School can teach broad skills, laboratories can test dangerous events, and forgiving jobs can support early independence. Apprenticeship combines useful deployment, customer revenue and real-world learning.

Foundation models can provide general competence, and deployment supplies the detailed knowledge required for dependable work. Investors can fund only a limited number of large robot laboratories; they cannot always pay for years of education across every machine and industry.

Founders should therefore design a robot’s first paid job alongside its autonomy plan.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?