Latest / Hardware Startups with Fexingo: Devices, Robotics, and Manufacturing Tech Companies / Inside the Industrial Robotics Startup That Raised $200 Million
Transcript
- Lucas: So last month, a robotics startup you've probably never heard of — Covariant — announced it had raised two hundred million dollars from investors including Lux Capital and the Canadian pension fund PSP. And the interesting thing isn't just the number. It's what they're trying to do. Luna: I've seen the headlines. Warehousing robotics, right? But there are dozens of those companies. Lucas: Right, but Covariant is attacking the hardest problem in warehouse automation: general-purpose picking. Most robots today are programmed for one specific task — picking a single type of box, or a single item. Covariant builds ai powered arms that can pick up anything: a shoe, a jar of peanut butter, a stuffed animal — things they've never seen before. Luna: That sounds like a software problem more than a hardware problem. What's the actual technology? Lucas: Exactly right. The hardware is fairly standard — a six-axis robotic arm from a supplier like Fanuc or Kuka. The magic is in the 'brain' — a deep neural network they call the Covariant Brain. It's trained on millions of real-world grasps from their installed base of robots in warehouses around the world. They use something called a vision transformer, which is a type of AI architecture originally designed for language, adapted for 3D spatial understanding. Luna: Wait — language AI adapted for picking up objects? How does that work? Lucas: So a vision transformer treats an image like a sentence — it breaks the image into patches, or 'tokens', and learns the relationships between them. For Covariant, those patches include depth information, texture, and even weight estimates. The network predicts a success probability for thousands of possible grasp points in real time. The robot chooses the one with the highest confidence and executes the pick. It's basically a decision engine that gets better the more data it sees. Luna: That data advantage is real. Every robot they deploy generates more training examples. But doesn't that mean the first customers get a worse product than later ones? Lucas: That's a fair critique, and it's a common problem in hardware-software startups. But Covariant addresses it by offering a 'robotics as a service' model. You pay per pick — or per hour of operation — and they upgrade the software continuously. So early customers still benefit from the network effect because the model improves across all units. The hardware is amortised over time. It's like a subscription for intelligence. Luna: Let's talk about the unit economics then. What's the cost per pick or per hour? And how does that compare to a human worker? Lucas: They haven't published exact numbers, but industry estimates suggest Covariant's system costs around eight to ten dollars per hour per robot, including the robot arm, sensors, compute, and software. A warehouse worker in the US earns about fifteen to twenty dollars per hour including benefits, and they can only work so many hours. A robot runs twenty-four hours a day. So in high-throughput environments, the cost per unit moved can be significantly lower. Luna: But humans are flexible. They can pack boxes, handle returns, do quality checks. A robot can only pick. So isn't this a partial replacement at best? Lucas: It's definitely not a total replacement. But consider this: in a typical e-commerce warehouse, picking accounts for roughly fifty-five percent of labor costs. It's the most repetitive, physically demanding task. If you can automate that, you're cutting your largest expense. And Covariant is already expanding beyond picking — they're working on packing and kitting, which are the next logical steps. Luna: What about competition? You've got Amazon Robotics with years of deployment, and they're developing their own AI. Plus there are companies like RightHand Robotics and Osaro in the same space. Lucas: Amazon is definitely the 800-pound gorilla. They've deployed over half a million robots in their fulfillment centers. But Amazon's robots are mostly custom-built for their own supply chain. They don't sell them to other companies. Covariant is going after the third-party logistics market — the companies that run warehouses for multiple retailers. That's a huge addressable market, and it's fragmented. No dominant player has emerged yet. Luna: So what's the biggest risk for Covariant? Is it technology, or is it go to market? Lucas: I'd say it's go to market. Selling to warehouse operators is a long, relationship-driven sales cycle. These companies are risk-averse. A single robot failure can halt a sortation line. Covariant needs to prove reliability at scale, and that takes time. Their two hundred million dollars gives them a runway of maybe three to four years. But they need to convert pilot customers into multi-site deployments to generate the recurring revenue that justifies their valuation. Luna: And their valuation? Reports say the round valued them at just under two billion dollars. That's a pretty high multiple on what I assume is still relatively modest revenue. Lucas: It's likely a bet on the platform. Investors are pricing in the idea that if Covariant's brain becomes the de facto standard for robotic manipulation, the value is enormous. Think of it as the operating system for robots. But it's a high-risk bet. If a competitor like Google or OpenAI decides to enter this space with their own vision transformers, the moat narrows. Luna: Speaking of big tech, Google had a project called 'RT-2' which was a vision language action model for robotics. That's pretty similar conceptually. Do you see that as a threat? Lucas: RT-2 was impressive — it used web data to teach robots general skills. But Google has a history of prototyping and then not commercializing. Covariant's advantage is that they have a real product in real warehouses generating real data. That data is the key differentiator. You can't just download a model; you need the feedback loop from actual pick failures and successes. Luna: So the thesis is: first-mover advantage in data network effects, plus a recurring revenue model, plus a huge market. That's the pitch. But what could go wrong? Lucas: A couple of things. First, the technology might hit a plateau. General-purpose manipulation is an ai complete problem — there's no guarantee that scaling up data will continue to improve performance. Second, the hardware costs could stay too high to compete with human labor in low-wage countries. And third, there's the classic hardware startup trap: over-promising and under-delivering. If a major customer has a bad experience, it could set the company back years. Luna: I want to zoom out a bit. This episode is about hardware startups. What's your takeaway for someone thinking about building a hardware company today? Lucas: The lesson from Covariant is: your differentiation should be in the software layer, not the hardware. Build on commodity hardware if you can, and focus on the intelligence. That lowers your capital requirements and lets you iterate faster. Also, think about a business model that generates recurring revenue — hardware is a one-time sale, but software as a service creates a long-term relationship. Luna: And the flip side? What if you have a hardware innovation? Lucas: Then you better have a really clear patent strategy and a path to manufacturing at scale. Hardware is hard. But the companies that succeed — like Tesla in cars or DJI in drones — they own the vertical stack: hardware, software, and sometimes even the manufacturing process. Covariant doesn't own the robot arm, and that's okay because their value is in the brain. Luna: So it's a strategy of 'software eats hardware'. Lucas: Exactly. And I think we'll see more of that. The next big hardware startup might look nothing like a hardware company. It might be a software company that happens to need a robot arm to deliver its service. Luna: Alright, I'm curious to see where Covariant is a year from now. Thanks, Lucas. Lucas: Always a good conversation, Luna. See you next time.