DreamTraj Predicts 6-DoF Trajectories from Single RGB Images
July 31, 2026
DreamTraj predicts 6-DoF object trajectories using unrendered video diffusion latents and a new MOVE dataset of 5,038 egocentric trajectories. It functions without requiring depth, CAD models, or video inputs at inference time.
HOW THIS AFFECTS YOU
●
builderYou can implement robotic manipulation control loops using only RGB input and language instructions.
●
researcherThis framework bypasses the need for expensive perception pipelines in trajectory prediction.