AFFORD-X project page

Meta-World scenes

Handle Points in 3D

The functional term scores a grasp by its distance d to the nearest handle point: 1 within 3 cm, falling to 0 at 8 cm. The handle points come from one of two sources. Pick an object, switch the source on the same scene, and compare how the grasps on the other part score.

drag to orbit · scroll to zoom · right-drag to pan
RGB render of the scene
RGB pass
Depth render of the same frame, near is yellow
Depth pass

Grasp candidates

Distance to the nearest handle point and the functional score it gives, for both sources.

#true partgate r·a·c d, geometricf, geometric d, SAM3f, SAM3 recorded rollout

How depth is rendered

No camera is added to the model. A MuJoCo free camera is placed in Python and one offscreen renderer draws the same simulator state twice, a colour pass and then a depth pass in metres along the optical axis. The depth is noiseless and has no dropouts.

From pixel to point

Pixel (u, v) with depth z is scaled by the intrinsics and moved into the world by the camera pose, as in affordx.rendering.camera.lift. The scene shows every third pixel within 1 m of the object.

Which source the numbers use