TL;DR WarmPrior replaces the standard zero-mean Gaussian source with a prior centered on the previous action forecast. This yields straighter flows and lower integration error for more accurate action generation with fewer steps in flow-matching models.
A flow-matching policy starts from Gaussian noise.It generates actions conditioned on the current observation.
WarmPrior improves success across Robomimic and MimicGen, including single-step generation.
| Task | NFE 1 | NFE 3 | NFE 9 | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Base | CM | WP-Past | WarmPrior | Base | CM | WP-Past | WarmPrior | Base | CM | WP-Past | WarmPrior | |
| Robomimic · state | ||||||||||||
| Square-PH | 83.6 | 80.9 | 86.6 | 87.3+3.7 | 86.2 | 84.5 | 88.0 | 87.9+1.7 | 86.7 | 85.5 | 88.1 | 88.1+1.4 |
| Square-MH | 65.9 | 70.1 | 70.1 | 77.8+11.9 | 65.4 | 70.5 | 73.2 | 72.9+7.5 | 65.9 | 67.1 | 69.2 | 72.7+6.8 |
| Transport-PH | 36.8 | 38.5 | 39.8 | 47.6+10.8 | 39.0 | 37.9 | 44.0 | 49.1+10.1 | 34.1 | 34.8 | 36.2 | 43.3+9.2 |
| Transport-MH | 23.3 | 24.1 | 30.2 | 34.5+11.2 | 21.3 | 21.2 | 30.7 | 30.4+9.1 | 16.3 | 18.3 | 20.7 | 24.3+8.0 |
| Tool-Hang-PH | 77.7 | 72.7 | 78.2 | 81.9+4.2 | 72.3 | 67.9 | 75.1 | 75.8+3.5 | 79.4 | 80.0 | 80.6 | 82.8+3.4 |
| Robomimic · image | ||||||||||||
| Square-PH | 88.7 | 88.3 | 89.3 | 89.1+0.4 | 87.7 | 88.9 | 89.2 | 89.6+1.9 | 86.9 | 87.2 | 88.2 | 88.7+1.7 |
| Square-MH | 72.4 | 73.3 | 77.6 | 75.1+2.7 | 73.8 | 73.2 | 77.9 | 77.1+3.2 | 76.1 | 72.1 | 78.0 | 77.8+1.7 |
| Transport-PH | 91.3 | 92.0 | 93.4 | 93.7+2.4 | 92.1 | 91.8 | 93.9 | 94.9+2.9 | 92.8 | 92.0 | 94.5 | 94.3+1.6 |
| Transport-MH | 74.3 | 72.9 | 78.6 | 79.7+5.4 | 73.8 | 72.9 | 80.0 | 80.7+6.9 | 74.8 | 76.0 | 79.7 | 79.8+4.9 |
| Tool-Hang-PH | 41.3 | 37.9 | 38.9 | 54.0+12.7 | 36.9 | 39.2 | 38.4 | 50.7+13.8 | 43.7 | 50.0 | 45.8 | 56.3+12.6 |
| MimicGen · image | ||||||||||||
| Stack | 21.3 | 21.8 | 22.4 | 28.7+7.4 | 21.3 | 20.9 | 23.7 | 30.7+9.4 | 21.4 | 22.8 | 22.8 | 31.6+10.2 |
| Coffee | 16.2 | 16.2 | 20.4 | 29.4+13.2 | 23.3 | 26.5 | 24.1 | 33.4+10.1 | 26.8 | 29.0 | 29.6 | 34.7+7.9 |
| Threading | 12.5 | 14.0 | 15.6 | 18.0+5.5 | 16.3 | 16.2 | 16.6 | 22.0+5.7 | 13.8 | 15.1 | 15.5 | 20.9+7.1 |
NFE = generation steps. Bold marks the best result within each task and NFE.Base uses a Gaussian prior. CM matches WarmPrior’s prediction horizon. WP-Past uses previously executed actions.
Swipe to compare all NFE settings →
The measured trajectories are straighter with WarmPrior.
A single generation step also produces actions closer to the reference solution.
5.24→4.10
4.41→3.38
Square-MH · selected policy trajectories · ten most curved paths per panel; shared axes.
Diagnostics: Square-MH images · normalized by target-action energy · ×10⁻² · lower is better.
WarmPrior improves GR00T N1.5 on a real Franka robot.
76%→87%
33%→45%
43%→67%
15%→41%
Mean success rate · Gaussian prior → WarmPrior.
WarmPrior also complements real-time chunking with π₀.₅.
DSRL-WP learns a bounded offset from the previous forecast.
The pretrained WarmPrior policy stays frozen and converts these prior samples into actions.
On Square and Transport, DSRL-WP learns faster and more stably than DSRL-NA and DSRL-SAC.
See the paper for full experiments and implementation details.
@article{kang2026warmprior,
title = {WarmPrior: Straightening Flow-Matching Policies with Temporal Priors},
author = {Kang, Sinjae and Kim, Chanyoung and Wang, Kaixin and Zhao, Li and Lee, Kimin},
journal = {arXiv preprint arXiv:2605.13959},
year = {2026}
}