Movement Trend Guidance lifts 3D diffusion policies without explicit trajectories — 72% vs 49% on real robots

dalian-university-of-technology · hf · 2026-09-21

Researchers propose Movement Trend Guidance, giving 3D diffusion policies foresight without explicit plans. The policy learns a compact latent of interaction evolution from a short observation history, supervised at training time by sparse future gripper states; at inference only the latent remains as future-oriented conditioning, plus a gated FiLM branch at the UNet bottleneck.

Adding just 3.52% parameters to DP3, it reaches 62.8% vs 56.1% on 50-task RoboTwin2.0 mixed training, 71.93% vs 37.08% on LIBERO-40, and 72.0% vs 49.0% on five real-robot tasks — showing diffusion policies benefit substantially from anticipating where an interaction is heading.

Original post →

More from Embodied

Embodied channel →