Robotics preprint adds 3.5% more parameters to improve diffusion-policy success

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to build on yet: no code, no weights, no evaluation dates, and every gain is measured against the authors' own DP3 baseline, not an external one. The idea itself is cheap — a few extra parameters and sparse future gripper states — so worth trying on your own policy if you have the data.

An arXiv preprint describes Movement Trend Guidance, an add-on to DP3, a diffusion policy for robot control. The add-on learns a compact numerical summary of how an interaction is developing from a short history of observations. During training, sparse future gripper positions supervise that summary; at inference it conditions the action generator alongside the current observation. It adds 3.52% more parameters to DP3. Against its own DP3 baseline, success rose from 56.1% to 62.8% on 50-task RoboTwin2.0, from 37.08% to 71.93% on LIBERO-40, and from 49.0% to 72.0% across five real-robot tasks. No evaluation dates are given.

Key facts

  • ·Movement Trend Guidance adds 3.52% more parameters to the DP3 policy it is built on. source
  • ·On 50-task RoboTwin2.0 mixed training, the method reaches 62.8% versus 56.1% for DP3. source
  • ·On LIBERO-40 it reports 71.93% versus 37.08% for DP3. source
  • ·Across five real-robot tasks it reports 72.0% versus 49.0% for DP3. source
  • ·Evaluations cover RoboTwin2.0, LIBERO-40 and DexArt plus real-robot tasks, with no evaluation dates given. source
  • ·Sparse future gripper states supervise the learned latent during training; at inference only the latent and current observation condition action generation. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire