A preprint reports that RoMAN-Flow’s One-Step policy generated a block of robot actions in 81.5 milliseconds on the LIBERO-Long benchmark, compared with approximately 697 milliseconds for its teacher—an 8.55× speedup. The student recorded 93.0% success, against 92.2% for the teacher.
The paper evaluates whether RoMAN-Flow can make autoregressive normalizing-flow policies practical for likelihood-based offline reinforcement learning in robotic manipulation, while reducing sampling overhead during training and inference.
The system has three moving parts
The training pipeline has three stages: imitation learning, NF-IQL post-training and one-step distillation. It also relabels rewards at the trajectory level to help assign credit when rewards are sparse.
NF-IQL updates the policy using action chunks already in the offline data, applying advantage-weighted exact-likelihood optimization without sampling actions from the current autoregressive policy.
For deployment, the post-trained autoregressive teacher is distilled into a bidirectional student that generates an entire action chunk in one forward pass while matching the teacher’s intermediate reverse states and final chunk.
Results varied by test
On MetaWorld-MT50, the reported unweighted mean success rate was 81.1% for NF-IQL, compared with 72.8% for imitation learning. The One-Step student reached 78.5%, and NF-IQL was 3.0 percentage points above π0 + Flow-SDE under the aligned protocol.
On LIBERO, NF-IQL averaged 95.3% success versus 93.5% for imitation learning. The paper reports a 6.6-percentage-point difference on LIBERO-Long; One-Step averaged 93.7% overall and was highest on LIBERO-Long.
On RoboMimic MH, NF-IQL scored 100% ± 0 on Lift, 96% ± 3 on Can and 85% ± 3 on Square. One-Step scored 100% ± 0, 95% ± 2 and 80% ± 3, respectively, compared with 91% ± 4, 96% ± 2 and 68% ± 5 for SERNF (TD3+BC). These figures are means ± standard deviations over four training seeds, with 100 evaluation rollouts per seed.
The physical-robot test was smaller
The real-robot evaluation covered four tasks: Pick Beaker, Pick Cylinder, Place Beaker and Put Beaker on Balance. Across them, NF-IQL averaged 81.5% success, compared with 57.3% for imitation learning—a reported difference of 24.2 percentage points.
The One-Step student averaged 69.5% across the real-world tasks, while NF-IQL was reported as 9.7 percentage points above π0.5.
In a RoboMimic Square-MH capacity test, S/B/L actors ranging from 33.8 million to 466.2 million parameters reached 77%–78% success. The XL configuration, with 685.5 million parameters, reached 85%, the highest reported score in that ablation.
What the results leave open
The findings are bounded by the evaluation design. Policies were tested from fixed checkpoints using episode-level success rates, and the real-world evidence covered four named tasks. The results therefore describe performance under those protocols rather than showing how the method would generalize beyond them.
Uncertainty was not reported uniformly. For RoboMimic, the paper gives mean ± standard deviation over four training seeds and 100 rollouts per seed; the LIBERO-Long latency is described as approximate, with no variability estimate given.
One-step distillation also involved a performance trade-off: the student averaged 93.7% on LIBERO versus 95.3% for NF-IQL and 69.5% on the real-world tasks versus 81.5% for NF-IQL, even as it was slightly ahead of its teacher on LIBERO-Long, at 93.0% versus 92.2%.
Paper data and sources
Original title: RoMAN-Flow: Taming Autoregressive Normalizing Flows for Offline Reinforcement Learning in Robotic Manipulation
Authors: Shaoxuan Wang, Guangting Zheng, Rui Huang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text