Learning Foresight without Explicit Trajectories for 3D Diffusion Policies
Movement Trend Guidance adds latent foresight to 3D diffusion policies, improving DP3 across RoboTwin2.0, LIBERO-40, and real-robot tasks with only 3.52% more parameters.
The method learns a compact latent representation of interaction evolution from short observation histories, supervised by sparse future gripper states during training and retained as conditioning at inference. A gated FiLM branch applies the latent at the UNet bottleneck. It improves DP3 to 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.
- Latent foresight conditioning avoids explicit trajectory planning
- Only 3.52% parameter overhead over DP3
- LIBERO-40 success jumps from 37.08% to 71.93%
From the original article
1 min read at huggingface.co
3D diffusion policies are strong at generating geometrically grounded actions from current observations, but successful manipulation requires not only knowing what motion is feasible now, but also anticipating where the interaction is heading. Existing policies largely leave such foresight to emerge implicitly from action learning.
By Zhongbo Zhang, Zaibin Zhang, Yifan Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.