Bifrost

Bifrost builds simulation and synthetic data software for physical AI, offering two products. Stardust generates photorealistic, multi-modal synthetic data: users describe a 3D scenario, typically in Python or through an agent driving the API, and receive fully labeled, sensor-matched output without requiring 3D expertise. The system supports controllable environmental variation (fog, mist, rain, snow, high winds, night operations), sensor perturbation (noise, intrinsics, perspective shifts), and real2sim pipelines that reconstruct field failures as 3D assets so teams can patch, retrain and measure the effect. Output spans RGB, infrared, depth and segmentation modalities in registration, with radar listed as forthcoming, across robotics, maritime, aerial, space and off-road domains, drawing on a library of more than 1,000 objects.

Manifold is a robot policy evaluation platform that runs a single harness across simulator benchmarks including LIBERO, LIBERO-Plus, RoboCasa, RoboMimic, CALVIN, SIMPLER, RLBench, RoboTwin 2.0, RoboLab and RoboMemory, as well as customer-defined scenarios, and supports simulators such as NVIDIA Isaac Sim, Unreal Engine, MuJoCo, ManiSkill and Genesis. It evaluates VLA and embodied AI policies of the OpenVLA, GR00T, pi-0 and Octo classes, sharding runs across GPUs so that a 1,000-episode evaluation that would take roughly eight hours returns in about 30 minutes. Features include decomposition of tasks into atomic subgoals, user-defined grading rubrics, rollout QA, and agentic failure analysis that clusters rollout failures by scenario, object, sensor, lighting and trajectory, ranks them by performance impact, and supports natural-language querying.

At the time of its 2024 Series A the company positioned itself as a generative 3D data platform addressing the shortage of domain-specific training data for physical AI, citing use cases such as autonomous boats patrolling shipping lanes, industrial robots assembling new parts, and drones inspecting electrical grids.

Founding story

TechCrunch reports that Charles Wong and Aravind Kandiah founded Bifrost in 2020. Wong had worked on AI perception models for self-driving cars at nuTonomy, an MIT spinout in autonomous vehicles and mobile robots, while Kandiah had built a medical AI system detecting early signs of blindness and diabetic retinopathy. Kandiah said the pair concluded that AI and robotics require enormous volumes of high-quality data and started the company with the goal of solving that data problem for physical-world applications.

Business model

Bifrost sells software to AI, robotics and autonomy teams rather than data-labeling services, generating revenue through an annual subscription model. Customers access controllable 3D simulation and synthetic data generation through a Python library and API, with Stardust offered via demos and Manifold distributed through an early-access waitlist. The company also works directly with industrial customers to codify real-world work into custom simulation tasks and success rubrics.

Revenue is generated through an annual subscription to the software platform.

Traction

TechCrunch reported 22 staff across the U.S. and Singapore and $13.7 million in total capital raised as of the October 2024 Series A, with the product then in closed beta with select heavy-industry partners and revenue from annual subscriptions. The company cites collaboration with NASA JPL on data generation engines for Moon and Mars exploration, partnerships with major U.S. government organizations and heavy-industry enterprises, and states it works in production with demanding teams. Founders Charles Wong and Aravind Kandiah are described as Forbes 30 Under 30 honorees, and the team is drawn from organizations including NVIDIA, Industrial Light and Magic, Ubisoft, Google and Meta.

Latest developments

The company now markets two products intended to close what it calls a recursive improvement loop: Stardust for synthetic data generation and Manifold for policy evaluation. Manifold is offered on an early-access basis, adding automated sim evaluation and grading, atomic subgoal scoring, custom rubrics, massively parallel GPU execution and agentic failure clustering, with additional benchmarks and embodiments described as in progress and radar listed as a coming sensor modality in Stardust.