Research map / Reinforcement learning / Visual and representation RL
Hierarchical Equivariant Policy via Frame Transfer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Hierarchical Equivariant Policy (HEP), which connects a high-level keypose predictor and a low-level equivariant diffusion policy through a translation-only Frame Transfer interface, achieving SO(2)/T(3) equivariance for sample-efficient robotic manipulation.
Problem
Hierarchical visuomotor policy learning from limited demonstrations is hampered by rigid interfaces between high- and low-level agents and by ignoring domain symmetries in robotic manipulation.
Approach
A Frame Transfer interface passes only a 3D translation (keypose) from the high-level agent, which is used as the coordinate origin for the low-level trajectory generator, so the low-level can refine trajectories flexibly; equivariant networks are used at both levels so the whole hierarchy is T(3)×SO(2) equivariant.
Key results
- Open-loop simulation: HEP averages 88% vs 78% for Chained Diffuser and 56% for 3D Diffuser Actor across 30 RLBench tasks, outperforming on 28 of 30.
- Closed-loop on 10 tasks: HEP averages 79% vs 57% for EquiDiff (+23%).
- Real-world open-loop success: HEP 0.8/0.85/0.9 across three tasks vs 0.3/0.2/0.4 for Chained Diffuser.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HEP rather than Chained Diffuser | RLBench open-loop policy · RLBench 30 tasks, open-loop | mean success rate (%): 78 -> 88; +10% | Helped |
| Use HEP rather than EquiDiff | RLBench closed-loop policy · RLBench 10 selected tasks, closed-loop | mean success rate (%): 57 -> 79; +23% | Helped |
| Use HEP rather than HDP | RLBench open-loop policy · RLBench 7 tasks | mean success rate (%): 74 -> 94; +20% | Helped |
| Use HEP rather than Chained Diffuser | real-world manipulation policy · Pot Cleaning | success rate: 0.3 -> 0.8 | Helped |
| Use HEP rather than Chained Diffuser | real-world manipulation policy · Blocks to Drawer | success rate: 0.2 -> 0.85 | Helped |
| Use HEP rather than Chained Diffuser | real-world manipulation policy · Blocks Stacking | success rate: 0.4 -> 0.9 | Helped |
| Train HEP with one demonstration and test on unseen object poses | HEP · Pick-place, one-shot | success rate: 80% | Helped |
| Add color variation | Block Stacking policy · Block Stacking, color variation | success rate: 0 -> 0.9 | Helped |
| Add color variation and distractor objects | Block Stacking policy · Block Stacking, color and distractor objects | success rate: 0 -> 0.6 | Helped |
| Add equivariant structure (ablation) | HEP · Six-task ablation suite | mean success rate: 0.70 -> 0.94 | Helped |
| Add Frame Transfer (ablation) | HEP · Six-task ablation suite | mean success rate: 0.78 -> 0.94 | Helped |
| Add stacked voxel encoder (ablation) | HEP · Six-task ablation suite | mean success rate: 0.84 -> 0.94 | Helped |
Surprising findings
- Removing equivariance hurts more than removing Frame Transfer (ablation mean drops 0.94 to 0.70 without equivariance vs 0.78 without FT).
- HEP achieves 98% on Take Money at earlier checkpoints but fails at the final checkpoint, attributed to overfitting.
Limitations
- Experiments focus on tabletop manipulation.
- No memory mechanism; poor performance on history-dependent tasks such as pot cleaning loops in closed-loop real-world runs.
- Frame Transfer only handles translation; rotational keyposes are not covered.
Experimental setup
Datasets: RLBench demonstrations
Benchmarks: RLBench (30 tasks) real-world UR5e manipulation tasks
Metrics: success rate
Baselines: 3D Diffuser Actor Chained Diffuser EquiDiff HDP
Builds on: diffusion policy Chained Diffuser Equivariant Diffusion Policy stacked voxel representation escnn
Closest papers by meaning
- $\mathrm{SO}(2)$-Equivariant Reinforcement Learning (ICLR 2022 · similarity 0.89)
- ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY (ICLR 2025 · similarity 0.88)
- Hierarchical Diffusion Policy for Kinematics-Aware Multi-Task Robotic Manipulation (CVPR 2024 · similarity 0.87)
- Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.87)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.86)
- Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution (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.