Physically accurate simulation and data to make robot learning transfer from simulation to the real world
Uncharted Dynamics has announced its venture to build physics infrastructure for embodied AI. Its high-fidelity, multi-body dynamics solver is designed to support the company’s expansion across the North American robotics market, and aims to reshape how robot learning models transfer from simulation to the real world.
The ongoing hypothesis, redefined
The prevailing hypothesis in AI is that scale alone drives capability: more data, larger models, and more computational power. In embodied AI, that recipe carries a hidden prerequisite. Scale only compounds into capability when the underlying data is physically consistent. This is rarely the case in robotics, where success is seldom determined by what a camera sees.
Instead, success in robotics is decided at contact: how force propagates through a grasp, how friction evolves across a surface, and how deformable objects respond to load. These signals are typically absent, oversimplified, or systematically approximated in simulation pipelines, where the resulting errors compound downstream.
"We’re not arguing against scale," said Zhewen He, CEO of Uncharted Dynamics. "We’re arguing that scale requires a physically accurate foundation first. Otherwise, more data just amplifies a systematic error."
The core of Uncharted Dynamics’ technology
At the core of Uncharted Dynamics’ platform are high-fidelity, multi-body dynamics solvers engineered for the exact failure modes of mainstream simulation tools: deformation, soft-rigid coupling, soft contact, and contact-rich manipulation.
Many off-the-shelf engines handle rigid-body kinematics well. However, they tend to lose fidelity exactly where it matters most: during dexterous manipulation, soft-object handling, and long-horizon assembly tasks.
A physics solver is not a renderer. Where a graphics engine computes what something looks like, a physics solver explains why a given outcome occurs: how force moves through a system, how contact is made, and how material states change. For robots, this causal chain is precisely what a model needs in order to learn reliably.
The company calls this output Physics-Augmented Data: datasets enriched with causal physical labels that cameras cannot observe and sensors cannot collect at scale. These include full 6-DoF contact wrenches, deformation feedback, friction characteristics, and material responses.
Building infrastructure into real-world models
Today the focus is on data and simulation infrastructure. The longer-term ambition is far larger.
As robotics moves toward systems that can predict, plan, and reason about outcomes, the quality of their internal physics becomes the binding constraint. Uncharted Dynamics is positioned to solve the problems of that next generation, not only through data, but through the layer that makes grounded world models possible.
The solver, datasets, and evaluation infrastructure the company is developing today will become the foundation for a physics-native model stack, where simulation and model capabilities compound together.
"Some companies are building bodies. Some are training brains," He said. "We are building the physically reliable classrooms those brains have to grow in."
The team, based in Montreal, combines a background in computational neuroscience with deep expertise in physics simulation and large-scale systems. The result is researchers who understand data working alongside engineers who can model the physical world from the ground up.
As the field of robotics continues to evolve, Uncharted Dynamics aims to make embodied AI more reliable and more interpretable over the long term, by building the foundational technology on which future systems will be built.
About Uncharted Dynamics
Uncharted Dynamics is a foundational infrastructure company based in Montreal, Canada. The company develops high-fidelity, multi-body dynamics solvers and physics-grounded data systems that enable reliable learning for robotics and embodied AI.
VentureBeat newsroom and editorial staff were not involved in the creation of this content.