SyncWorld-Evaluation

Evaluation sets for yyuncong/SyncWorld, a robot world model that predicts future video from past frames plus end-effector actions.

set	source	tasks	episodes	camera views
evaluation_maniskill	ManiSkill	5	50	base_camera_rgb, render_camera_rgb
evaluation_libero	LIBERO	5	50	agentview_rgb, side_rgb

Each episode is a directory holding <view>/video.mp4 plus pose.pkl, with a sibling calibration/ directory containing the per-DoF calibration sweep the model conditions on:

evaluation_maniskill/
└── PushCube-v1/
    └── episode_0/
        └── camera_poses_0/
            ├── expert/                 # the episode
            │   ├── render_camera_rgb/video.mp4
            │   └── pose.pkl
            └── calibration/            # per-DoF sweep
                ├── render_camera_rgb/video.mp4
                └── pose.pkl

Usage

hf download yyuncong/SyncWorld-Evaluation --repo-type dataset --local-dir ./SyncWorld-Evaluation

See the SyncWorld README for how to run an evaluation.

Contents

Tabletop manipulation only: a robot arm, a table, and toy objects. No people appear in any frame.

License

OpenMDW-1.1, matching the SyncWorld model and the NVIDIA Cosmos-Framework it derives from. The sets are renders from ManiSkill and LIBERO; refer to those projects for their own terms.