Research

Research Highlights

DreamVu Research is bridging the knowledge gap for physical AI, creating the high-fidelity datasets required for the next generation of humanoid robots and embodied agents.

SABER: A Scalable Action-Based Embodied Dataset for Real-World VLA Adaptation.

The first high-fidelity retail robotics action dataset built from natural human behavior, not teleoperation.

The Impact

Domain-specific robot deployment is fundamentally a data problem. SABER demonstrates that human video — systematically captured and retargeted — is a scalable foundation for robot adaptation, achieving 2.19X improvement over baselines.

View Full Research Hugging Face

• 44.8K Training Samples
• 100+ Hours Captured
• 91% Fridge Task Success Rate

PRISM: Unifying physical AI knowledge across space, physics, and embodied action.

Bridging the reasoning-action gap for VLMs in real-world environments.

The Impact

PRISM provides the structural bridge between visual understanding and physical execution, reducing average error rates by 66.6% across embodied reasoning tasks.

View Full Research Hugging Face

• 66.6% Error reduction
• 270K Samples
• 5× Reasoning Performance Gain