Research map / Reinforcement learning / Visual and representation RL
HiPolicy: Hierarchical Multi-Frequency Action Chunking for Policy Learning
Jiyao Zhang, Zimu Han, Junhan Wang, Xionghao Wu, Shihong Lin, Jinzhou Li, Hongwei Fan, Ruihai Wu, Dongjiang Li, Hao Dong
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 HiPolicy hierarchical multi-frequency chunking and entropy-guided execution | DP · RoboTwin 1.0/2.0 (all 21 tasks) | success rate (%): 37 -> 60 | Improved |
| add HiPolicy | DP3 · RoboTwin 1.0/2.0 (all 21 tasks) | success rate (%): 41 -> 59 | Improved |
| add HiPolicy | DP · RoboTwin precise tasks | success rate (%): 22 -> 45 | Improved |
| add HiPolicy | DP3 · Pick apple messy | success rate (%): 10 -> 3 | Hurt |
| use hierarchical condition instead of fixed low-frequency condition (ablation) | DP+HiPolicy · RoboTwin (12 tasks) | average success rate (%): 58 -> 60 | Improved |
| use hierarchical condition instead of fixed high-frequency condition (ablation) | DP+HiPolicy · RoboTwin (12 tasks) | average success rate (%): 49 -> 60 | Improved |
| add hierarchical feature fusion attention (ablation) | DP+HiPolicy · RoboTwin (12 tasks) | average success rate (%): 54 -> 60 | Improved |
| add multi-frequency hierarchical actions (ablation) | DP+HiPolicy · RoboTwin (12 tasks) | average success rate (%): 37 -> 60 | Improved |
| add entropy-guided execution (ablation) | DP+HiPolicy · RoboTwin 2.0 (5 precise tasks) | execution steps: 139 -> 100 | Improved |
| add entropy-guided execution (ablation) | DP+HiPolicy · RoboTwin 2.0 (5 precise tasks) | success rate (%): 45 -> 41 | Mixed |
| increase entropy sampling N from 1 to 100 (ablation) | DP+HiPolicy · RoboTwin 2.0 (Scan object etc.) | inference time (ms): 105.5 -> 107.5 | No clear effect |
| raise entropy thresholds toward high-frequency execution (ablation) | DP+HiPolicy · Scan object | execution steps: 45 -> 47 SR; 108 -> 143 steps | Mixed |
Closest papers by meaning
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.88)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.88)
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.88)
- Real-Time Execution of Action Chunking Flow Policies (NeurIPS 2025 · similarity 0.88)
- PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.87)
- Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling (ICLR 2025 · similarity 0.87)
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.
