From Passive Video to Editable Experience: Physically Grounded Experience Synthesis for Embodied Intelligence addresses the "embodiment gap" that prevents robots from learning directly from the billions of human manipulation videos available online. Because human and robot morphologies, kinematics, and action spaces differ, raw video cannot be directly used as training data. Author Jia Luo introduces Pegasus, a framework that translates human demonstrations into robot-learnable data by converting video into structured, physically grounded graphs rather than relying on pixel-level imitation.
Bridging the Embodiment Gap
Pegasus treats robot data generation as a knowledge transfer problem rather than a hardware collection task. The framework extracts a Task Graph from human videos, which captures sub-goals, objects, and spatial relationships. This graph is then transformed through an Affordance Graph and a Constraint Graph into a Robot Planning Graph. By focusing on these intermediate representations, the system separates embodiment-invariant task semantics from embodiment-specific execution requirements, such as joint limits and collision constraints.
Hierarchical Affordance Latent Space
To enable generalization, Pegasus uses a hierarchical Affordance Latent space. Instead of relying on specific object identities—which limits a robot to objects it has seen before—the model organizes representations by object states, affordances (such as graspability or pourability), and task intentions. This allows the system to perform tasks on unseen objects by recognizing their functional properties. For example, the model can determine how to handle a watermelon by identifying it as "graspable, heavy, and fragile," even if it has never encountered that specific object during training.
Physics-Constrained Verification
Visual fidelity does not guarantee that a robot can actually perform a movement. Pegasus includes a closed-loop physics verifier that checks generated trajectories against kinematic feasibility, collision constraints, and joint limits. If a generated plan fails these checks, the system produces error signals that are fed back into the planning module to iteratively refine the trajectory. Experiments show this process reduces inverse-kinematics errors by 93.5% and collision rates by 90.9%, with 94.1% of trajectories becoming executable within five iterations.
Performance and Policy Learning
In experiments using benchmarks like GTEA Gaze+ and EPIC-KITCHENS-100, Pegasus-generated data improved downstream robot policy performance. When training models like OpenVLA and ACT, adding Pegasus-generated data to real-robot demonstrations resulted in an 11.3% improvement in success rates compared to using real-robot data alone. Furthermore, the framework’s graph-conditioned generation outperformed prompt-only video generation methods by 12.3% in task correctness and 15.7% in executability, suggesting that structured task representation is a primary driver of successful cross-embodiment transfer.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!