Research map / Reinforcement learning / Visual and representation RL
Hierarchical Policy Learning via Spectral Decomposition
Shuxin Cao, Liquan Wang, Walker Byrnes, Yiye Chen, Yilun Du, Animesh Garg
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 language to the low-frequency predictor (ablation) | CSP · LIBERO-90 K=16 | success rate (%): 22.2 -> 91.6 | Improved |
| add language to the low-frequency predictor (ablation) | CSP · LIBERO-10 K=16 | success rate (%): 64.5 -> 81.0 | Improved |
| add low-frequency conditioning to the high-frequency predictor (ablation) | CSP · LIBERO-10 K=32 | success rate (%): 51.5 -> 75.0 | Improved |
| predict low frequency before high frequency (not reversed) (ablation) | CSP · LIBERO-90 K=16 | success rate (%): 49.6 -> 91.6 | Improved |
| predict low frequency before high frequency (not reversed) (ablation) | CSP · LIBERO-10 K=16 | success rate (%): 35.5 -> 81.0 | Improved |
| add hierarchical coarse-to-fine conditioning to spectral diffusion (ablation) | Frequency Diffusion · LIBERO-90 Set A large noise | success rate (%): 73.0 -> 87.5 | Improved |
| add causal frequency modeling over no-hierarchy (ablation) | No Hierarchy · LIBERO-90 Set B large noise | success rate (%): 72.5 -> 80.0 | Improved |
| add hierarchy on Stack under teleoperation noise (ablation) | Frequency Diffusion · Stack (noise) | success rate (%): 81.3 -> 75.0 | Hurt |
| add hierarchy with no noise on Set B (ablation) | No Hierarchy · LIBERO-90 Set B no noise | success rate (%): 96.0 -> 94.0 | No clear effect |
| vary low-frequency cutoff 20%-40% (ablation) | CSP · LIBERO / MimicGen | success rate (%) | No clear effect |
| use CSP over Action Binning at chunk size 64 | Action Binning · MimicGen mean (chunk size 64) | success rate (%): 31.8 -> 50.0 | Improved |
| use CSP over ACT at longer chunk size 64 | ACT · MimicGen mean (chunk size 64) | success rate (%): 9.0 -> 50.0 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BAKU — BAKU: An Efficient Transformer for Multi-Task Policy Learning (NeurIPS 2024) · better on 2 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
- From Noise to Intent: Anchoring Generative VLA Policies with Residual Bridges (ICML 2026 · similarity 0.87)
- Hierarchical Equivariant Policy via Frame Transfer (ICML 2025 · similarity 0.87)
- Learning High-Frequency Continuous Action Chunks in Latent Space (ICML 2026 · similarity 0.87)
- HiPolicy: Hierarchical Multi-Frequency Action Chunking for Policy Learning (ECCV 2026 · similarity 0.87)
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.87)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.86)
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.
