Robot costs are collapsing and capability is compounding. What's scarce is the company that turns the pieces into a product the market will adopt.
In the spring of 1961, a two-ton hydraulic arm called Unimate went to work at a GM plant outside Trenton, New Jersey. It pulled hot die castings out of a press. It was the first industrial robot ever deployed, and it came with a promise. Machines would take over dull, dirty, and dangerous work. That didn’t happen in the very early days of robotics because, in part, Unimate and its peers were very expensive relative to the labor they aimed to replace.
By 2015, a welding robot cost roughly half a year's salary for the welder it was replacing. And yet robots still had made a relatively small dent in the manufacturing workforce. It turns out that the expensive part was making the robot useful: months of custom integration that turned a $25,000 arm into a six-figure deployment, once tooling and integration were added. While hardware had gotten cheaper over the years, integration never did.
Over the last decade, two things have shifted in robotics. The first is foundation model robot policies that make it faster and less expensive to train and deploy robots. At Physical Intelligence, where we're investors, solutions powered by PI’s models take days to set up rather than months. And second, the EV industry spent two decades scaling motors, batteries, and sensors and in doing so accidentally built a lower-cost robotics supply chain along the way. The average industrial robot cost about $46,000 in 2010. By 2022, it was $23,000.
At the same time, labor costs have continued to climb. A U.S. manufacturing worker costs $46.30 an hour fully loaded. That number compounds at 3-4% a year and it has never gone down. Compare the robot cost curve to the manufacturing labor cost curve, and they cross right about now. The cost of a robot hour is now falling through the cost of a labor hour on that same task.
.png)
Provided that foundation model robot policies keep improving and hardware continues to get less expensive, the trend toward lower cost and more capable robots will only continue. And as it does, it will make more economic sense for companies across a greater number of industries to consider deploying more robots.
Components don't make markets
However, the models and hardware are just components used to create a robot solution for an end customer.
Components have never made a market. The companies that use components to build end solutions for customers make markets. These are the assemblers, and they are mission critical for the next generation of robotics to take off.
Microprocessors didn't put a computer on every desk; Compaq and Microsoft did. Commodity servers didn't give corporations the cloud; Amazon did by turning racks into something a CIO could sign for. Component makers create the possibility. Assemblers create the market and in turn reap much of the economic reward.
The assembler's work is unglamorous. They must pick a task narrow enough to master. Fit the robot into the workflow the customer already runs instead of asking them to rebuild around it. Ensure the solution is safe. Win over the plant manager, the foreman, the crew standing next to the machine. Answer the phone when it fails at 2 a.m. This is hard work.
We’ve seen success in assembly
When we led DroneDeploy's Series A in 2015, drone hardware was becoming a commodity. What was missing was software that, combined with a drone, could turn aerial photos into something a superintendent could actually act on. DroneDeploy built just that, and drones quietly became the first robots deployed at scale across many verticals including construction, energy, and public safety.
Bedrock Robotics uses the same logic on today's stack. Construction can't hire enough operators, and no contractor will scrap a working fleet for an unproven machine. So Bedrock retrofits autonomy onto the excavators contractors already own, fits inside the safety rules that already govern a jobsite, and sells against empty seats. In both cases, the components mattered. But the components alone would have sold nothing.
What assemblers have to get right
Across our robotics portfolio and the hundreds of other robotics companies we've met, three challenges come up again and again for assemblers:
1. Unit economics until autonomy arrives.
The economics look ugly before the robot is truly autonomous. Early deployments lean on teleoperation and on-site babysitting, which means you're paying humans to supervise the machine that was supposed to replace labor. That's how you kickstart the data flywheel.
Waymo and Ambi Robotics both built their edge this way, using teleoperated and supervised deployments to generate the data that steadily reduced how much human oversight each one needed. But it has to be rigorously planned, or you'll run out of cash.
Know which milestones unlock your next raise before you need it: interventions per hour trending down, deployments per teleoperator rising, one lighthouse customer expanding rather than five pilots idling.
One way to improve the math is to avoid buying hardware. Bedrock sits atop the iron contractors already own, which takes the largest capital cost out of the model and turns a hardware business into something much closer to software economics. If your vertical has installed machines with empty seats, start there.
2. The third shift is a system, not a robot.
The headline pitch we hear constantly: with robots, a customer can run a third shift and get to 24-hour operation. Then you discover that while the robot is ready for the third shift, nothing else about the customer's operation is. No maintenance staff overnight, no materials handler feeding the line.
One solution that companies claim they will pursue to address this is the automation of all processes with robots. Theoretically, this can work provided a sufficient level of accuracy and reliability can be reached. But chain ten processes together at 95% reliability each, a level most individual robotic systems can hit today, and the combined line's uptime falls to 59%. If overnight operation is part of your value proposition, you have to design for the whole system from the beginning.
3. Edge compute is a workflow decision.
A robot often can't wait on a round trip to a data center. Latency means an unusable machine. Foundation model labs are solving this with action chunking, where the policy predicts a sequence of moves per inference call (a technique that originated in Stanford's ALOHA/ACT research and has since become standard practice across the field) and with distilled on-device models that let a modest onboard chip keep up with a big policy.
But onboard compute draws more power, which creates its own operational challenges. Assemblers must treat tethering, battery swaps, and charging windows as workflow decisions and make them before the pilot, not during it.
Seeking assemblers
Unimate's promise that machines would take over mundane work took sixty-five years to come due. For most of that time, integration cost was the reason: hardware got cheap, but making a robot useful on a specific job never did. That's finally changing.
The market needs the assemblers that will develop and deploy magical next-generation robot-powered solutions. Building these companies is an important part of the work in robotics over the next decade. At Emergence, we aim to partner with the founders who are doing this important work. If you're assembling, we'd like to talk.
Kevin Spain is a General Partner at Emergence Capital. Emergence is an investor in Physical Intelligence, Bedrock Robotics, and DroneDeploy.
