Research map / Reinforcement learning / Offline RL
Flow-Based Single-Step Completion for Efficient and Expressive Policy Learning
Prajwal Koirala, Cody Fleming
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add single-step completion policy (SSCQL) in place of diffusion actor DQL | DDPG+BC offline RL · D4RL MuJoCo average | normalized score: 87.9 -> 87.9 | No clear effect |
| replace DQL's 5-step denoising actor with one-step SSCQL | offline RL actor · training and inference time | training time: ~8 hours -> ~16 mins | Improved |
| use one-step SSCQL instead of 5-step DQL denoising | offline RL actor · inference latency | inference time (ms): 1.27 -> 0.27 | Improved |
| use SSCQL instead of DQL/CAC in fully online RL | online RL · Online RL average | normalized score: 67.7 -> 90.53 | Improved |
| use SSCQL instead of DQL in offline-to-online finetuning on Hopper-ME | offline-to-online finetuning · Hopper-ME | normalized score: 71.7 -> 111.27 | Improved |
| increase SSCQL offline/online steps from 100K+100K to 250K+250K (ablation) | SSCQL finetuning · Halfcheetah-ME | normalized score: 97.32 -> 110.68 | Improved |
| add single-step completion policy in place of Diffusion Policy for behavior cloning | BC · RoboMimic/Push-T average | success rate: 0.88 -> 0.87 | No clear effect |
| SSCP vs Diffusion Policy on Toolhang | BC · Toolhang | success rate: 0.87 -> 0.64 | Hurt |
| add shared-network hierarchy distillation (GC-SSCP) in place of hierarchical HIQL | goal-conditioned offline RL · OGBench average | success rate: 50.94 -> 54.44 | Improved |
| GC-SSCP vs HIQL on antmaze-large-stitch | goal-conditioned offline RL · antmaze-large-stitch-v0 | success rate: 67 -> 40 | Hurt |
| add completion loss (alpha_2 > 0) to flow matching loss (ablation) | SSCQL actor · D4RL MuJoCo (Hopper, Walker2d medium) | normalized score | Improved |
| add completion loss on HalfCheetah-Medium (ablation) | SSCQL actor · HalfCheetah-medium | normalized score | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- HIQL — HIQL: Offline Goal-Conditioned RL with Latent States as Actions (NeurIPS 2023) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Flow-Based Policy for Online Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- DeFlow: Decoupling Behavior-Prior Modeling and Value Maximization for Offline Policy Extraction (NeurIPS 2026 · similarity 0.90)
- Flow Q-Learning (ICML 2025 · similarity 0.90)
- Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.89)
- Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport (NeurIPS 2026 · similarity 0.89)
- Flow Matching for Offline Reinforcement Learning with Discrete Actions (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
