I remember the meeting like it was yesterday. We were three months into the drone project, and the burn rate was terrifying. The damn things just couldn't get it right. We were trying to build an autonomous warehouse inspection system, and the drones kept misidentifying pallets, bumping into racks, and generally behaving like expensive, drunken hummingbirds. My co-founder looked at me, his face pale, and said, "Sahin, I think we have to pull the plug."
He was right, logically. The data was awful. Our top-tier AI models, the ones that were supposed to be revolutionary, were failing over 60% of the time. We had burned through nearly half a million dollars, and all we had to show for it was a collection of slightly dented drones and a blooper reel of warehouse crashes. We were on the verge of giving up. I’ve had two successful exits and invested in over 200 companies, including giants like Anthropic and OpenAI, but this little drone company was about to break me.
This is the side of innovation nobody likes to talk about. It’s not a clean, linear path. It’s a messy, brutal fight against reality. And in the world of physical AI, where software meets the unforgiving laws of physics, the fight is even tougher.
The Brutal Reality of Hardware
People who’ve only worked in software don’t always get it. In software, you can spin up a thousand instances, run A/B tests for pennies, and deploy a patch in minutes. When you’re building a robot, your iteration cycle is measured in days or weeks, not hours. A failed test doesn’t just mean a line of red text in a log; it means a broken gearbox, a fried circuit board, or a 200-pound machine tipping over. The financial cost of a failed hardware test is orders of magnitude higher than a software bug. You can't just git revert a shattered sensor array.
We see the slick marketing videos from companies, and it all looks so easy. But behind every seamless robot demo are thousands of hours of failures. I saw this firsthand with a company I invested in that was developing warehouse robots. They were trying to automate the pick-and-place process for a major e-commerce player. The initial promise was a 99.9% accuracy rate. For the first six months, they were stuck at 70%. That’s a 30% error rate in a high-volume warehouse. It was a disaster. Every day, the CEO would send me a report, and every day my heart would sink. They were trying everything—new grippers, different camera angles, bigger datasets. Nothing worked.
The pressure from the client was immense. The board was getting nervous. We had a very serious conversation about whether to pivot or shut down. But the founding team had grit. They went back to first principles. They realized their simulation environment didn't accurately model the physics of how different items—a soft bag of chips versus a rigid box of soap—behaved when gripped. They spent two solid months just rebuilding their simulation. It was a huge gamble, pouring resources into a non-customer-facing problem. But it paid off. When they deployed the new software, accuracy jumped to 98% overnight. A week later, it was 99.5%. They saved the company by focusing on the unglamorous, difficult physics of the problem. It’s a powerful lesson: in robotics, simulation is not a nice-to-have; it's the bedrock of success.
Humanoids are Coming, But Not Tomorrow
This brings me to the current fascination with humanoid robots. I’m an investor in Figure AI, and I’m blown away by what Brett and his team are doing. The videos of their robot making coffee are incredible. It’s a monumental engineering achievement. Similarly, what Tesla is doing with Optimus is pushing the entire field forward. The progress is undeniable.
But I have to be the voice of reason here. We are still a long, long way from seeing humanoid robots deployed at scale in complex, unstructured environments. Walking is hard. Manipulation is harder. Doing both at the same time while navigating a chaotic human world is exponentially harder. The coffee-making demo is a huge step, but it’s a controlled environment. The robot knows exactly where the machine is, where the cup is, how the button works. It’s not figuring that out on the fly.
Consider the sheer complexity of a seemingly simple task like stocking a grocery shelf. The robot needs to identify thousands of different SKUs, handle delicate items like eggs and heavy items like a gallon of milk, navigate around unpredictable shoppers, and deal with imperfectly stocked shelves. The number of edge cases is nearly infinite. This is where the gap between a controlled demo and real-world value becomes a chasm. The last 1% of the problem takes 99% of the effort.
My prediction? We’ll see specialized robots dominate for the next decade. The economics just make more sense. A robot designed to do one thing, like unload a truck or move boxes, can be built far more cheaply and reliably than a general-purpose humanoid. The ROI is clearer. Why build a robot that can do everything when you only need it to do one thing, 10,000 times a day? Companies like Boston Dynamics understood this. They wowed us with the acrobatics of Atlas, but their commercial success came from Spot, a quadruped with a clear, specific job. The future, for at least the medium term, belongs to the specialists.
The Unseen Revolution: Surgical Robots
While everyone is watching the humanoid race, a quieter revolution is happening in the operating room. Surgical robotics isn't new—the da Vinci system has been around for over two decades—but the integration of advanced AI is turning these systems into something else entirely. I’ve looked at a few startups in this space, and the potential is staggering.
Think about it. A surgeon’s hands, no matter how steady, have natural tremors. A robot has none. An AI can analyze thousands of surgical videos to identify the most efficient way to make an incision or the safest path to navigate around a delicate nerve. It can provide real-time feedback to the surgeon, highlighting risks they might not see. This isn't science fiction; it's happening now.
We’re moving from robots as remote-controlled hands to robots as active, intelligent partners. One company I’m tracking is using AI to create a 3D map of the patient’s anatomy from a CT scan, and then using that map to guide the surgical robot with a level of precision a human could never achieve. This isn't about replacing surgeons. It's about augmenting them, giving them superpowers. It reduces errors, shortens recovery times, and makes complex procedures accessible to more patients. This is where physical AI is already having a massive, life-changing impact, and it’s only going to grow. The FDA approvals are the main hurdle, but the technology is ready.
Investing in the Physical World
As an investor, looking at robotics companies is a different beast than evaluating a SaaS startup. The metrics are different. You can't just look at Monthly Recurring Revenue (MRR) and Customer Acquisition Cost (CAC). You have to look at the bill of materials (BOM), the supply chain, the manufacturing plan, and the team's expertise in mechanical and electrical engineering. A brilliant AI researcher with no hardware experience is a huge red flag.
I look for teams that are obsessed with the problem, not the solution. The teams that win are the ones that have an almost painful intimacy with the customer's workflow. They've spent hundreds of hours in the warehouse, on the factory floor, or in the operating room. They understand the messy reality of the environment their robot will live in. They aren't just building a cool piece of tech; they are solving a real, costly problem.
Another key is the approach to data. Many teams think they just need a massive dataset. But the best teams I've seen focus on creating a high-quality, targeted dataset and a powerful simulation environment. They understand that the data flywheel is what will ultimately create a defensible moat. Their robot isn't just performing a task; it's constantly learning from every action, every success, and every failure, making the entire system smarter with each passing day.
So, What Happened to the Drones?
We didn't give up. In that dark moment, when we were ready to throw in the towel, one of our junior engineers, a kid barely two years out of college, spoke up. He had a crazy idea. He noticed that the drones weren't just failing at identifying objects; their flight paths were erratic, especially in the narrow aisles between the 30-foot-tall warehouse racks. He theorized that the drone's own propellers were creating air turbulence that interfered with its navigation sensors.
It sounded absurd. The senior engineers dismissed it. It was a classic case of "not invented here" syndrome, combined with the understandable skepticism towards a junior team member challenging the core assumptions of the project. But we were out of options. I told them to give him a week. Let him try it. He spent that week building a new computational fluid dynamics model and integrating it into the flight controller. It was a hack, a complete long shot.
He uploaded the new code to Drone #7, the most beat-up one in our fleet. We all stood back. He armed the motors, and it lifted off. It was different. It was stable. It flew down the aisle, its camera locked onto the pallets. It scanned the first one. The second. The tenth. It completed the entire aisle without a single error. The silence in the room was deafening. Then, a cheer erupted.
That kid saved our company. He didn't invent a new AI model. He just looked at the whole problem—the software, the hardware, the physics—and saw the connection everyone else had missed. That’s the lesson. The future of robotics isn’t just about bigger models or more data. It’s about a holistic understanding of how software and the physical world interact. It’s about the grit to keep going when every piece of data tells you to stop. And sometimes, it’s about listening to the crazy idea from the kid in the corner. That’s where the real breakthroughs happen. Not in a pristine lab, but in the messy, frustrating, beautiful collision of code and reality.
Frequently Asked Questions
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
How can I apply this thinking to my own situation?
Start by identifying the core principle behind the opinion, not the specific example. Then ask yourself: does this principle apply to my context? If yes, test it in a small, low-risk way before going all in.