At the 2023 CVPR Workshop on Autonomous Driving, GL of Momenta took the stage to deliver a clear message: the path to true full autonomy cannot rely on traditional approaches; instead, it requires a data-driven flywheel that lets algorithms self-improve from massive amounts of real-world driving data. Since its inception, Momenta has set its sights on scalable Level 4 autonomy—making every car on the road capable of driving itself anywhere, anytime—with a firm safety threshold: the system must be at least ten times safer than the average human driver.
The core challenge lies in the millions of long-tail scenarios that any robust autonomous system must handle. GL estimated that solving these edge cases and validating the solutions could demand over 100 billion kilometers of driving data. A few hundred or even a thousand test vehicles could never gather enough. This led to two key insights. First, rule-based engineering is a dead end: you can hand-craft hundreds of rules with a hundred engineers, but you cannot hand-craft millions of rules with millions of engineers; data-driven approaches are the only scalable way. Second, the only viable path to L4 is to deploy the software onto millions of mass-produced vehicles and intelligently harvest data from them. Only by tapping into production fleets can we approach the data volumes needed.
Based on these insights, Momenta built a data-driven flywheel with three pillars: data-driven software (using machine learning across perception, localization, and planning instead of hand-written rules), mass-production data collection from vehicles produced by multiple automotive partners, and a closed-loop automation toolchain that smartly filters, processes, labels, and trains models, enabling automatic iteration at scale. GL noted that the company has already accumulated 150 petabytes of historical data, with algorithmic iterations continuously supported by these production vehicles.
In the technical deep-dive, GL showcased Momenta's "mapless" solution. Traditional autonomous driving heavily relies on high-definition (HD) maps, which are expensive, slow to update, and geographically limited. Momenta, by contrast, relies only on standard-definition (SD) maps—the kind used in consumer navigation—as a rough prior. The system comprises three key modules. Perception takes SD maps, multi-camera inputs (lidar and radar are optional), and outputs a local map of surrounding objects and landmarks using an algorithm called DDLD (Data-Driven Landmark Detection). Localization fuses IMU, wheel speed, and GNSS (possibly with RTK) via another data-driven method, DDPF (Data-Driven Pose Fusion), which replaces conventional Kalman filters with a learned model to output precise ego-pose in the body frame while also obtaining a coarse global pose to align with the SD route. Finally, the planner receives the local map, ego-pose, and navigation route, then generates driving decisions and trajectory.
This architecture does raise some eyebrows. It still needs a coarse SD map and a rough global pose—acknowledging that pure mapless driving is not yet fully realized. Critics might argue that relying solely on real-time perception without a precise prior could falter in extremely complex or degraded environments. Moreover, while the data flywheel is conceptually powerful, its effectiveness hinges on the quality and diversity of the collected data; a fleet biased toward simple driving scenarios might not expose the system to rare, safety-critical events. Nevertheless, GL’s presentation made it clear that Momenta is betting on closing the loop at massive scale: letting the data shape every component, from perception to planning, thereby sidestepping the brittleness of rules and the cost of HD maps.
In sum, Momenta’s CVPR talk can be read as both a technical roadmap and a manifesto for an engineering philosophy: shift the focus from hand-engineering individual algorithms to building a self-sustaining data ecosystem where millions of cars encounter millions of edge cases, and automated toolchains turn those encounters into millions of model updates. Whether this flywheel will ultimately deliver full autonomy remains to be seen, but its underlying logic—meeting long-tail complexity with mass-produced data—offers a compelling and scalable framework for the entire industry.
