Xavier Chi says his company, Mbodi, ran its robot for eight hours straight and it succeeded 99.6% of the time.
Every large tech company chasing robotics has bet on the moonshot: a foundation model trained on enough data to do anything, the way ChatGPT did for text. Chi's bet is the opposite — skip the moonshot, deploy narrower AI into warehouses now, and let the resulting production data do what a research lab never can.
"It's kind of like a fact. We haven't seen a single successful robotic software company yet."
Chi co-founded Mbodi with a former Google teammate, neither of them a robotics specialist by training, won ABB Robotics' AI startup challenge out of a couple hundred entrants, and once taught a robotic arm to pour a Coca-Cola live on the TechCrunch Disrupt Startup Battlefield stage.
I listened to the full interview so you can skip it. 32 minutes of audio, 11 minutes of reading.
Here are the 7 lessons that matter.
👤 Guest: Xavier Chi, co-founder of Mbodi, which builds AI software that lets industrial and collaborative robots learn tasks from natural language
🎙️ Host: Isabelle Johannesen, who hosts Build Mode for TechCrunch
📰 Published: 10 September 2026 on YouTube (TechCrunch)
🔴 YouTube | 🟣 Apple Podcasts | ⏱️ 32 min | ✅ Time saved: 23 min
Key Takeaways
No robotics software company has yet generated meaningful revenue at scale, in Chi's view
He calls it close to a fact, not an opinion, after watching well-funded peers fail to get past a handful of deployed robots
The reason is customization, not technology
Every past deployment required bespoke engineering per site, which does not survive a VC-backed company's balance sheet
Generative AI is supposed to let the robot figure out the task in the moment, instead of being custom-programmed for it
That is Chi's whole bet on why this generation of robotics companies scales where the last one didn't
A recent eight-hour, single test run came back 99.6% successful, which he offers as the reliability bar a customer actually needs
Winning ABB's AI startup challenge, not a pitch competition, was the unlock that got Mbodi into large industrial accounts
The process ran several months and included an ABB engineer visiting Mbodi's office to test the product live
He does not want a humanoid. He wants a wheeled or four-legged robot, because the shape doesn't have to match a human's
Hardware is getting cheaper fast, so Chi expects software, not the robot itself, to capture most of the value going forward
1. Winning ABB's Challenge
Chi credited one relationship above all others for turning Mbodi from an idea into a company with large industrial customers.
ABB Robotics is the largest robot manufacturer in the world, and Mbodi won its AI startup challenge in late 2024 out of "a couple hundred companies"
The evaluation was not a pitch deck exercise. ABB sent an ecosystem manager to Mbodi's New York office to test the product, over several months and multiple rounds, before signing
The partnership works because the two companies don't compete. "They are amazing at making hardware and we're really good at making software," Chi said, and ABB brings its own large customers, particularly in consumer packaged goods
Chi's advice to founders picking which competitions are worth their time: judge the evaluation process and the caliber of who shows up, not the size of the prize
2. Why Robotics Stalls at Scale
Asked why robotics software companies keep failing to grow, Chi laid out what he called close to a fact: no one in the category has generated real revenue at scale yet.
He described a company founded by top robotics researchers with the best vision-picking model in the industry, unnamed, that hit the same wall every company before it did
The default scaling path is partnering with a systems integrator, who bundles the robotics software into a much larger warehouse-automation sale. But robotics is "only a very small piece" of that sale, and after two or three years of work a project might have five robots deployed
The alternative — deploying directly — runs into the same problem from the other side. Every customer needs customization, and "every time people put in customization into a solution it just not scalable"
The economics do not survive venture funding. "The balance sheet look crazy" under either path, Chi said, which is why he thinks a lot of robotics companies have run out of money rather than a lack of demand
3. AI Kills Customization
Chi's case for why this generation of companies is different rests on one distinction.
He splits the problem in two: the systems-integration layer isn't changing, but the customization layer is
Traditional automation required bespoke engineering for every deployment. Generative AI, in Chi's telling, lets the robot or the software figure out what needs to be done in the moment, instead of that work being hand-coded in advance
That is the whole argument for why this generation might scale where the last one didn't: less integration work per site, not a smarter robot arm
4. Proof: 8 Hrs, 99.6% Success
Reliability, Chi said, is the single most important thing in robotics, because a stopped line is not a minor defect — it is money lost every minute the line is down.
"We did a recent test run with our system. Eight hours straight from beginning to end. 99.6% success. So that's actually quite good."
Asked how that compares with a human doing the same job, Chi said there is no real benchmark: "Human also messed up. There's really no metric for like how often human mess up, but human do make a mistake as well"
His pitch to customers rests on two claims rather than one: the robot costs less, and it runs non-stop at a reliability he says matches a human's
He gave a related anecdote: one of his customers, one of the largest pharmaceutical companies in the world, was overhiring by 20% in its warehouse because workers kept failing to show up, and then lost staff again when a rival opened a warehouse next door and hired them away
5. Bootstrap, Then a Seed Round
Mbodi's fundraising path ran through a small pre-seed, an accelerator, and then a seed round closed after Y Combinator's spring 2025 batch — last June or July, by Chi's account.
Chi and his co-founder Sebastian are both Google alumni, not robotics specialists by background. Sebastian triple-majored at UPenn and studied in the university's GRASP Lab; Chi built a non-AI drawing robot in college
The two met as Google teammates and worked together for a little over two years, during which, Chi said, "we always talk about robots from time to time." Around 2024 they decided the moment was right to quit and start the company
Winning the ABB challenge came before the seed round and reshaped the pitch. It gave Mbodi direct exposure to industrial customers' actual problems, which Chi says is what ultimately convinces a VC the team can execute
One of the company's engineers came from Bell Labs. Mbodi is currently hiring robotics engineers and AI engineers who can orchestrate agents
6. The Impossible Triangle
Chi described robotics hardware economics as an "impossible triangle" that companies have had to solve all at once rather than trading off.
Customers weigh speed, cost and flexibility/capability together, and a company has to maximize all three rather than pick one
Industrial robot arms are expensive, and that cost has historically been a limiting factor on deployment
Chi said he has watched hardware get cheaper on a recent trip to China, where he was struck by the pace of manufacturing iteration, and expects the arms themselves to become closer to a commodity
That is why he thinks software, not the robot, will capture the most value going forward — hardware costs stop being the binding constraint once they keep falling
7. Consumer Robots Stay Toys
Asked where consumer robotics goes next, Chi drew a line between what he thinks is coming soon and what he still considers a long way off.
He does not believe a general-purpose home robot that does chores is close. The gap between an 80–90% success rate in a research paper and a warehouse deploying that same model ten times in a row — where a 10–20% failure rate compounds — is, in his telling, the central unsolved problem in the field
He personally would want a robot in his home, but not a humanoid one. "I always feel like they don't have to look like a human," citing warehouses where a wheeled or even four-legged design works fine
He pointed to a small interactive robot from Hugging Face running an MCP server as an early example of the category he expects to grow: fun, toy-like consumer robots rather than dishwasher-unloading humanoids
He also referenced Amazon's reported acquisition of a company he called "FA Robotics," which he said made a similarly small, interactive robot — a claim this summary could not independently confirm
On the industry landscape, he named the incumbent robot makers his software plugs into: ABB, Fanuc, Kuka and Universal Robots, alongside the systems integrators who assemble full warehouse solutions
Bonus Insights
The famous Battlefield Coke demo wasn't originally scripted to open the can. Chi said the decision to actually pour it, rather than leave the can closed, was made the same day as the pitch — "why not make Isabelle crazy?"
Chi decided as a college freshman, after watching Silicon Valley, that he would one day compete on the Startup Battlefield stage himself — and said the competition became an informal "stamp of approval" milestone for some founders as much as a fundraising step
The Battlefield stage produced advisors, not just attention. Chi said investors and advisors approached the team directly after their pitch
Asked about VCs' shifting appetite, Chi said there is now more investor interest in real deployed robots than in pure foundation-model research bets, compared with a few years ago when ChatGPT-style excitement dominated
Chi believes robots should free humans for more creative work rather than repetitive motion — "I believe humans should do more creative work instead of just repeating the same motion" — and brought up, half in jest, that both he and the host enjoy singing
The episode's throughline is that Chi is betting Mbodi's future on narrowing the gap between research-lab demos and warehouse floors through deployment and data, not on a bigger foundation model.
Products, Companies & Tools Mentioned
Mbodi (Chi's company: AI software that lets industrial and collaborative robots learn tasks from natural language and run in production)
ABB Robotics (The world's largest robot manufacturer. Its AI startup challenge, won by Mbodi in late 2024, became the company's main enterprise channel)
Fanuc, Kuka, and Universal Robots (The other major industrial robot makers Chi named alongside ABB as the hardware his software plugs into)
Y Combinator (Mbodi went through YC's spring 2025 batch, and closed its seed round shortly after)
Amazon and Kiva Systems (The host's example of large-company automation: Amazon's 2012 acquisition of Kiva Systems produced the shelf-to-station robots now common in its warehouses)
Hugging Face (Maker of a small interactive consumer robot, running an MCP server, that Chi cited as an early sign of where consumer robotics is heading)
TechCrunch Disrupt (Site of the Startup Battlefield stage where Mbodi's Coke-pouring demo took place, and a formative childhood ambition for Chi)
Bell Labs (Former employer of one of Mbodi's engineers, cited by Chi as evidence of the team's research depth)
University of Pennsylvania's GRASP Lab (Where co-founder Sebastian studied robotics as a triple major, before the two met at Google)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: