ACWMBench Inference Dataset
This self-contained package provides initial RGB observations and aligned action/state trajectories for action-conditioned world-model evaluation.
Only the initial observation is included. No future RGB frames are provided;
predicting future visual outcomes is the responsibility of each model adapter.
Every episode starts at frame zero, with timestamp equal to frame_index / fps.
Action and state vectors retain their recorded values.
Use manifest.json for episode discovery, the JSON schemas for structure, and
shared/g02/kinematics.json for the mesh-free robot description. SHA-256
manifests support integrity checks.