Amazon Bought Another Robot Company. I've Seen This Movie Before.
A former Amazon robotics engineer breaks down why the Rivr acquisition faces the same cost and scale challenges that every robotics effort at Amazon eventually hits.
TL;DR
- Amazon acquired Rivr, an ETH Zurich spinout that builds a four-legged wheeled delivery robot for “last meter” doorstep delivery
- I helped build robots at Canvas Technology, which Amazon acquired in 2019, and watched the same cost pressures play out firsthand
- Our robot cost $12,000 to build. Amazon challenged us to hit $3,000. For an autonomous robot with computer vision, that was nearly impossible
- Rivr’s robot is even more complex. Four articulated legs, wheels, stair-climbing capability, a 60-pound payload. The bill of materials will be brutal
- The “driver assist” model doesn’t eliminate labor. You still need people loading, unloading, and babysitting the robot. That changes the math
- Amazon already tried this with Scout and shut it down. Different approach, same fundamental economics problem
Last week, Amazon acquired Rivr, a Swiss robotics startup that builds quadruped delivery robots. The robot walks on four legs (with wheels for flat ground), climbs stairs, carries packages to your doorstep, and navigates around trash cans and pedestrians using “General Physical AI.”
It’s a genuinely impressive piece of engineering. And I’m skeptical it will ever make economic sense at Amazon’s scale.
I say this as someone who has been inside the machine. I was an engineer at Canvas Technology, an autonomous robotics company that Amazon acquired in 2019 for over $100 million. I’ve lived through the exact cycle that Rivr is about to enter.
The $12,000 Problem
At Canvas, our robot cost roughly $12,000 to build. It was a warehouse navigation robot with computer vision, spatial AI, and SLAM (simultaneous localization and mapping). After the acquisition, Amazon challenged us to get the cost down to $3,000.
Think about that for a second. A fully autonomous robot with computer vision, running on hardware that includes GPUs powerful enough for real-time inference. Just the graphics card would eat half that budget today. Some of the cost was architecture decisions that made sense as a startup (we ran server racks for local intelligence) but didn’t translate well into Amazon’s environment. Even accounting for that, $3,000 was a target you could see from a distance but never quite reach.
Now look at Rivr’s robot. It has four articulated legs, each with its own motors and sensors. It has wheels. It carries 60+ pounds. It climbs stairs. It navigates outdoors in rain, snow, and uneven terrain. It has the same computer vision and AI requirements as our robot, plus the mechanical complexity of a quadruped.
How much does this cost to build? Rivr hasn’t disclosed that number, and I’d bet there’s a reason. I would be shocked if the bill of materials comes in under $15,000 per unit, and that’s being generous. A comparable quadruped from Chinese competitors like Unitree runs in that range, and they benefit from massive manufacturing scale that Rivr doesn’t have yet.
Robots Break. A Lot.
Nobody in the press releases mentions this. Robots operating in uncontrolled environments break constantly.
A warehouse is relatively predictable. Climate-controlled, flat floors, known obstacles. Even there, our robots needed regular maintenance. Sensors get dirty. Parts wear out. Electrostatic discharge fries components. Software encounters edge cases that simulation never predicted.
Now take that robot and put it outside. On sidewalks with cracks and puddles. Up stairs covered in ice. Through yards where sprinklers are running. Past dogs that want to investigate (or attack) this strange metal creature.
Every failure requires intervention. Either a remote operator takes over (Rivr maintains a 24/7 support team for exactly this), or someone physically retrieves the robot. Each intervention costs money and slows down deliveries. At Amazon’s volume of millions of packages per day, even a 1% failure rate means tens of thousands of daily interventions.
The “Driver Assist” Shell Game
Amazon is positioning Rivr as a tool that helps drivers, not one that replaces them. A driver pulls up to a neighborhood, deploys the robot to walk a package to one doorstep while the driver handles another delivery. Efficiency gains without eliminating jobs.
I don’t buy the framing. But even if we take it at face value, think about the operational reality.
Someone has to load the robot into the van. Someone has to unload it at each stop. Someone has to place the package in the robot’s compartment. Someone has to monitor it as it navigates to the door. Someone has to troubleshoot when it gets stuck on a step or confused by a garden gnome.
If you’ve seen how fast Amazon expects its delivery drivers to move, you know where the tension is. DSP drivers deliver hundreds of packages daily, with quotas sometimes hitting 400 per shift. These drivers are under enormous time pressure. Adding a robot to babysit doesn’t simplify their job. It complicates it.
So if you still need a human in the van, loading, unloading, and monitoring, where are the cost savings? You’ve added the capital expense of the robot, the maintenance costs, the remote support infrastructure, and you still have the driver.
The Amazon Robotics Playbook
Amazon has a pattern with robotics acquisitions, and it’s worth tracing.
Kiva Systems (2012). Acquired for $775 million. This one actually worked. Warehouse robots moving shelves to human pickers on flat, predictable floors. Amazon now runs over a million of these. But the environment is controlled, the task is narrow, and the robots don’t climb stairs.
Canvas Technology (2019). Acquired for $100M+. After the acquisition, Amazon fired all of Canvas’s existing customers and focused the technology internally. That’s the standard playbook. Buy the company, absorb the talent, redirect the mission.
Amazon Scout (2019-2022). Amazon’s own sidewalk delivery robot. Tested for three years in multiple cities. Discontinued because it “wasn’t meeting customers’ needs.” About 400 employees were reassigned.
Covariant (2024). Not a full acquisition, more of a talent grab. Amazon hired the founders and licensed the AI, likely structured to avoid regulatory scrutiny.
Rivr (2026). Here we are again.
The pattern is pretty clear. Amazon excels at warehouse robotics, where the environment is controlled and the task is repetitive. Every attempt to bring robots into uncontrolled, variable environments (sidewalks, doorsteps, stairs) has struggled or failed.
The Real Costs of Human Drivers
Look, there are genuine problems with the current model. Amazon’s delivery operation has real costs that don’t show up in simple wage calculations.
Nearly 1 in 5 Amazon DSP drivers were injured on the job in 2021, a 40% increase from 2020. Trips, slips, falls, strains, dog bites. Workers’ comp claims. Vehicle accidents. Damaged packages. The last mile accounts for 53% of total shipping cost while representing only the final leg of the journey. And there’s a nationwide shortage of 78,000 commercial drivers.
These are real problems. But the question isn’t whether the current system is expensive. It’s whether robots are cheaper. And when your robot has four legs, a dozen actuators, multiple cameras, a GPU, a 60-pound payload capacity, and needs to operate in rain, snow, heat, and up flights of stairs, I don’t see how the math works.
What I Think Is Actually Happening
Amazon isn’t stupid. They know the economics are challenging. I think the Rivr acquisition is really about three things.
Talent. Rivr spun out of ETH Zurich with genuine expertise in legged locomotion and reinforcement learning. That talent is valuable regardless of whether the delivery robot ever ships.
Data. Every deployment, even at pilot scale, generates training data for Amazon’s broader AI efforts. The delivery robot might fail commercially but succeed as a data engine.
Optionality. Battery costs drop. Compute gets cheaper. Manufacturing scales. Maybe in five years the economics flip. Amazon can afford to place small bets ($22M in prior funding, acquisition price undisclosed but likely modest) on that possibility.
Wall Street seems to agree that this isn’t a big deal. Amazon’s stock barely moved on the announcement.
The Bottom Line
I’ve stood in Amazon’s robotics labs. I’ve watched brilliant engineers struggle to hit cost targets that physics wouldn’t cooperate with. I’ve seen what happens when you take a robot that works beautifully in a demo and try to run it 10,000 times a day in the real world.
Rivr’s technology is real. The engineering is impressive. But impressive engineering and viable economics are different things. Until someone can show me a quadruped delivery robot that costs less per delivery than a person walking the same package to the same door, I’ll remain skeptical.
Amazon can afford to run this experiment. Just don’t expect it to change how your packages get delivered anytime soon.
David Kerr is the founder of Kerrberry Systems and a former robotics engineer at Canvas Technology (acquired by Amazon). He builds custom software for businesses that want a partner, not a vendor. Find him on LinkedIn or GitHub.