Learning Foresight without Explicit Trajectories for 3D Diffusion Policies
A small foresight latent gave 3D diffusion robot policies sharper manipulation results without adding explicit plans.
The paper introduces Movement Trend Guidance, which learns a compact representation of how an interaction is evolving from recent observations. During training, sparse future gripper states supervise that latent; at inference, the policy uses only the latent alongside the current observation. The method adds 3.52% parameters to DP3 and keeps its dense-action, receding-horizon setup. Reported gains include 62.8% vs. 56.1% on RoboTwin2.0 mixed training, 71.93% vs. 37.08% on LIBERO-40, and 72.0% vs. 49.0% on five real-robot tasks. HF Daily Papers' note
The paper introduces Movement Trend Guidance, which learns a compact representation of how an interaction is evolving from recent observations. During training, sparse future gripper states supervise that latent; at inference, the policy uses only the latent alongside the current observation. The method adds 3.52% parameters to DP3 and keeps its dense-action, receding-horizon setup. Reported gains include 62.8% vs. 56.1% on RoboTwin2.0 mixed training, 71.93% vs. 37.08% on LIBERO-40, and 72.0% vs. 49.0% on five real-robot tasks. HF Daily Papers' note
score 5