Pegasus Bridges Embodiment Gap for Robot Learning from Human Videos

A significant hurdle in embodied AI is the scarcity of suitable training data for robots, despite the abundance of human manipulation videos online. The "embodiment gap" prevents direct learning from these human-centric demonstrations. This paper introduces Pegasus, a novel, low-resource framework designed to bridge this gap. Pegasus transforms human video demonstrations into data that robots can directly learn from. It achieves this by first extracting a "Task Graph" from human videos, then converting it into a "Robot Planning Graph" via Affordance and Constraint Graphs. A hierarchical affordance latent space helps generalize beyond specific objects, and a closed-loop physics verifier filters out kinematically impossible or collision-prone robot actions. The framework was evaluated across various manipulation benchmarks and robot embodiments, demonstrating reliable cross-embodiment translation. This approach redefines robot data generation from a hardware-intensive collection problem to a scalable knowledge transfer challenge.