Before a single Bifrost robot rolls across Martian regolith, it has already failed - spectacularly, repeatedly - inside a photorealistic simulation that Charles Wong's company generated from scratch in under an hour. That's the pitch. That's also the product. And increasingly, that's what NASA JPL, the U.S. Air Force, and some of the largest industrial companies on the planet are paying for.
Wong is the CEO and Co-Founder of Bifrost AI, a San Francisco-based generative 3D data platform. The company builds synthetic worlds - accurate enough, varied enough, strange enough - to train perception models for robots operating where failure is catastrophic and real-world data collection is nearly impossible. Think lunar terrain. Ocean shipping lanes. Military logistics corridors.
"This approach is brutal. It costs millions, takes years, and proves nearly impossible to scale."- Charles Wong on traditional real-world AI data collection
What Bifrost replaces is the years-long slog of deploying robots across hundreds of locations, collecting footage, labeling it by hand, running quality checks, and repeating the entire loop every time conditions change. Wong's platform compresses that cycle into hours. AI engineers - not 3D simulation specialists - run it themselves. That last part is the competitive moat. Unlike Nvidia's Omniverse, Bifrost doesn't require a dedicated graphics team to operate.
A Singaporean Engineer Who Won a Hackathon in Bonn
Wong grew up in Singapore, attended NUS High School, then enrolled at the Singapore University of Technology and Design (SUTD) - a university co-established with MIT that selects for engineers who want to build, not just study. At SUTD, he joined the electric vehicle racing team, spending long hours in workshop bays building cars. He later said he would "love to spend all my time in a workshop." He still might, technically - his workshops just happen to run on servers now.
In 2017, Wong and a small team entered the UN Climate Change Conference hackathon in Bonn, Germany. They won. The concept: an autonomous vehicle system addressing urban emissions. It was early work that pointed toward something real - AI-driven physical systems - but the implementation tools didn't yet exist. He filed that away.
That same year, he took a gap year with SUTD classmate Aravind Kandiah (later his co-founder and CTO at Bifrost). They called it "unbounded experimentation" - an intentional break from structured education to figure out what they actually wanted to build. It wasn't a sabbatical. It was reconnaissance.
The NuTonomy Years: Where the Data Problem Became Obvious
Before founding Bifrost, Wong joined nuTonomy as an Autonomous Vehicle Engineer. NuTonomy was an MIT spinout working on self-driving cars and autonomous mobile robots - one of the early serious players in AV perception. Wong's job was building city-wide mobility simulations and testing vehicle deployment strategies. The work gave him an unusual vantage point: he could see exactly where the bottlenecks were, and they weren't the algorithms.
The bottleneck was data. Getting enough of it. Getting the right kinds. Getting it labeled correctly. Getting it to reflect edge cases that rarely appear in real-world fleets - night fog in Singapore, anomalous pedestrian behavior, sensor degradation at extreme temperatures. By the time Wong left nuTonomy and returned to SUTD to finish his degree, he knew exactly what he wanted to build. He built the first Bifrost prototype in 2019.
What Bifrost Actually Does (Explained Without Marketing Language)
An industrial robotics company wants to train a computer vision model to detect corrosion on offshore oil platforms. To do that properly, they need thousands of labeled images showing corroded pipes in varying weather, angles, lighting conditions, and camera configurations. In the real world, collecting that data takes years of deployment, safety protocols, manual annotation, and constant iteration. With Bifrost, an AI engineer spins up a simulated offshore platform, configures the sensor parameters, dials in weather variability, generates 50,000 labeled images, and starts training - in the same afternoon.
The platform serves domains where physical data is dangerous, expensive, or simply impossible to collect at scale: maritime collision avoidance (piracy scenarios you can't safely stage), off-world navigation (there are no Mars fleets to deploy), aerial drone inspection (electrical grid work that grounds personnel), and military logistics (scenarios you cannot run in civilian airspace).
Platform Capabilities
Bifrost's CI/CD pipeline for physical AI benchmarks robotics systems against thousands of varied scenarios daily - flagging failures before they happen in the real world. Teams can patch specific data gaps, balance underrepresented classes, and stress-test models against adversarial conditions, all without touching hardware. The reduction in iteration cycles is roughly 100x compared to pure real-world fleet-based training.
The Fundraise: Aerospace, Government, and the Sequoia Connection
In October 2024, Bifrost AI closed an $8M Series A led by Carbide Ventures, bringing total capital raised to $13.7M. The investor list reads like a who's-who of physical AI's strategic money: Airbus Ventures (aerospace), Peak XV Partners - the former Sequoia Capital India & Southeast Asia operation (global institutional), Wavemaker Partners (Southeast Asia tech), and Techstars. The presence of Airbus Ventures as a strategic backer is particularly notable - it signals Bifrost's ambitions in aerospace simulation well beyond commercial robotics.
Wong was direct about the use of the capital: commercializing across aerospace, maritime, manufacturing, and national security. Not pivoting - accelerating.
NASA, the U.S. Air Force, and Building for Extreme Environments
The marquee collaborations tell the product story better than any pitch deck. Bifrost is working with NASA's Jet Propulsion Laboratory on data generation engines for Moon and Mars exploration. The challenge: how do you train a rover's perception system to handle Martian terrain when you have almost no real labeled data from the surface? You simulate it - with physically accurate sensor models, realistic lighting at a 1.5AU solar distance, rock formation distributions derived from orbital surveys, and thousands of synthetic traversal scenarios generated overnight.
The U.S. Air Force engagement points in a similar direction: high-stakes, data-scarce environments where failure isn't an option and real-world training data is either classified, dangerous to collect, or physically impossible to gather at sufficient scale.
"Don't Forget the Mission! As a deep tech founder, every challenge you face is exacerbated by the fact that you are building in deep tech."- Charles Wong on the realities of deep tech founding
The Founder Behind the Platform
Wong's personality shows through his LinkedIn presence in an endearing way. When Bifrost gets press coverage, he posts that he needs to "pinch himself" - the reaction of someone who genuinely didn't expect the world to keep validating the bet they made. A veteran mentor once told him to stop and acknowledge success rather than perpetually pushing forward without recognition. He took it to heart. He talks about it publicly. For a deep tech founder building for government contracts and space agencies, he's surprisingly willing to share the vulnerable stuff.
He describes his obligations to his team with unusual seriousness - framing it not as standard employer responsibility but as a genuine debt. The people who joined Bifrost chose it over safer careers. Getting the outcome right is, in his framing, partly about them. That ethos shapes the company's culture.
He's also practical in the way that people who've spent time in real engineering environments tend to be. His GitHub repositories from 2017-2019 are still public - experiments in autonomous lane detection, driver behavior cloning, self-driving simulation controllers. The code is clean. He knows how the sausage is made, which makes him better at selling the machine that makes it.
What Comes Next
Bifrost's trajectory points toward becoming infrastructure - the simulation and data layer that physical AI teams build on top of, rather than a project-by-project service provider. The platform's expansion into CI/CD evaluation pipelines (continuous benchmarking of robotics AI, not just training data) suggests Wong is thinking about the full development lifecycle, not just the data generation step.
The company operates from San Francisco and Singapore, a deliberate dual presence that connects U.S. government contracts with Asia-Pacific's manufacturing and maritime industries. Japan is named as a growing market. The physical AI adoption curve there - from shipbuilding to semiconductor fabrication - maps exactly to Bifrost's domain verticals.
Forbes named Wong to their 30 Under 30 list. NASA is generating Moon data with his platform. The U.S. Air Force is a client. That's an unusual set of facts for someone who started by winning a UN hackathon about autonomous cars in Germany. The throughline is obvious in retrospect. It usually is.