Building scene

DIMO Motion Lab

512 key points and 6,000 Gaussians per body, each Gaussian skinned from its 6 nearest key points.

Motion latent space

Each body on the stage plays the motion at one point of this sheet. Pick a body, drag its point, or resample the set.

Body 1z = (0.00, 0.00)

Prompts

Sets the selected body to the motion a caption maps to.

Decoder weights

Illustrative simulation. A basis of motions blended by RBF weights over a 2-D sheet stands in for DIMO's 32-D latent and 8-layer decoder; key points ride their link and Gaussians are skinned from their six nearest key points with the paper's RBF weights. The mechanism is described in the method section. The UR5e plays the project's nine motion clips. UR5e meshes © Universal Robots.