Bifrost Stardust Synthetic Data & Manifold

Bifrost AI was founded in 2020 in Singapore by Singapore University of Technology and Design alumni Charles Wong and Aravind Kandiah. It tackles the time and cost of collecting and labeling real-world training data for perception and robotics AI, especially for rare or dangerous scenarios at sea, in the air, off-road or in space.

Stardust generates synthetic 3D environment datasets with pixel-perfect labels, scenario metadata and RGB, infrared, depth and segmentation outputs from thousands of 3D assets. Manifold, launched in 2026, evaluates robot policies at scale: one harness runs policies across simulator benchmarks on Isaac Sim, Unreal Engine, MuJoCo, ManiSkill and Genesis, grades rollouts against rubrics and clusters failures with AI agents. Users include robotics, defense, maritime and aerospace teams.

• Stardust synthetic data: Generates 3D scenes with pixel-perfect labels and scenario metadata.
• Multi-sensor outputs: Produces RGB, infrared, depth and segmentation data for perception models.
• Manifold evaluation: Runs robot policies across simulator benchmarks in one harness.
• Rubric grading: Grades rollouts with atomic subgoals and custom rubrics.
• Agentic failure analysis: Clusters failures, ranks them by impact and answers natural-language queries.
• Simulator support: Works with Isaac Sim, Unreal Engine, MuJoCo, ManiSkill and Genesis.

A Vietnamese drone or UAV company can train object detection on synthetic aerial and infrared imagery. A university robotics lab can benchmark manipulation policies on Manifold before testing on real arms. A shipbuilder or port operator can generate maritime scenes to train collision avoidance for autonomous vessels.

Teams can obtain labeled training data for rare scenarios in hours instead of running costly field collection, and the company claims training and testing can be accelerated 100x. Manifold reports running 1,000 evaluation episodes in 30 minutes instead of 8 hours through parallel GPU evaluation. Automatic failure clustering shows which errors matter most, reducing manual review of robot rollouts.

Physical AI for robots, drones and autonomous vessels needs far more training and test data than can be collected in the field, so synthetic data and simulation are becoming standard. Southeast Asian manufacturers, ports and research labs, including in Vietnam, are starting robotics and autonomy programs that face the same data shortage.

Founded in 2020 by SUTD alumni Charles Wong (CEO) and Aravind Kandiah (CTO). Backed by Sequoia, Lux Capital, Airbus Ventures, Wavemaker, Carbide, Techstars and Cap Vista, with a Series A announced in October 2024. Logos shown include NASA, Honda, NTT Data, ClassNK and ST Engineering.