Research map / Reinforcement learning / Visual and representation RL
Hierarchical Diffusion Policy for Kinematics-Aware Multi-Task Robotic Manipulation
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 Diffusion Policy (HDP), pairing a PerAct high-level next-best-pose agent with a low-level RK-Diffuser that generates joint-position trajectories distilled from end-effector pose diffusion via differentiable forward kinematics.
Problem
Multi-task language-guided robotic manipulation where next-best-pose agents rely on fixed motion planners that ignore task context, and end-effector pose diffusion policies suffer inverse-kinematics failures that violate robot kinematic constraints.
Approach
Factorise the policy into a high-level PerAct next-best end-effector pose predictor and a low-level goal-conditioned diffusion model (RK-Diffuser) that jointly learns end-effector pose and joint-position trajectory diffusion, distilling the accurate pose diffusion into joint space through differentiable forward kinematics and inpainting with the goal pose.
Key results
- HDP achieves 80.18% overall success rate across 11 RLBench tasks, versus 71.27% for PerAct + Diffuser and 57.72% for PerAct + Planner.
- RK-Diffuser reaches 94.55% overall success with ground-truth next-best poses, versus 83.45% for RKD-ResNet and 73.64% for joint diffusion.
- On a real Franka Panda, HDP achieves 100% success on opening an oven and 94% on sorting objects into a drawer, using 10 demonstrations per sub-task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HDP instead of PerAct + Diffuser | RLBench overall · RLBench | success rate: 71.27 -> 80.18 | Helped |
| use HDP instead of PerAct + Diffuser | RLBench open oven · RLBench: open oven | success rate: 18 -> 58 | Helped |
| use HDP instead of PerAct + Diffuser | RLBench open box · RLBench: open box | success rate: 82 -> 90 | Helped |
| use HDP instead of PerAct + Planner | RLBench take lid off saucepan · RLBench: take lid off saucepan | success rate: 100 -> 96 | Hurt |
| use HDP instead of PerAct + Planner | RLBench knife on board · RLBench: knife on board | success rate: 76 -> 72 | Hurt |
| replace RK-Diffuser with RRT (ablation) | low-level control with ground-truth next-best poses · RLBench overall | success rate: 94.55 -> 26.82 | Hurt |
| use end-effector pose diffusion with inverse kinematics instead of RK-Diffuser (ablation) | low-level control with ground-truth next-best poses · RLBench overall | success rate: 94.55 -> 67.18 | Hurt |
| add differentiable-kinematics distillation to joint-position diffusion (ablation) | low-level control with ground-truth next-best poses · RLBench overall | success rate: 73.64 -> 94.55 | Helped |
| remove depth and use pretrained ResNet50 features (ablation) | RK-Diffuser · RLBench overall | success rate: 94.55 -> 72.18 | Hurt |
| replace PointNet++ feature extraction with ResNet (ablation) | RK-Diffuser · RLBench overall | success rate: 94.55 -> 83.45 | Hurt |
| use HDP on the real robot | Franka Panda · Real-world open oven | success rate: 100% | Helped |
| use HDP on the real robot | Franka Panda · Real-world sorting objects into drawer | success rate: 94% | Helped |
Surprising findings
- Pose Diffusion suffers a 24.55% IK error rate, with invalid quaternions causing 75% of its failures.
- Joint position diffusion without last-step inpainting underperforms RK-Diffuser notably on open oven (74 vs 95 on pick up cup area, and 15 vs 100 on toilet seat up).
Limitations
- Error accumulation over long-horizon tasks may cause distribution shift and failure in behaviour cloning.
- Open microwave task performance is low because the high-level agent's final pose distribution is highly diverse.
- Real-robot evaluation uses only two tasks.
Experimental setup
Datasets: RLBench demonstrations real-robot demonstrations (Franka Panda)
Benchmarks: RLBench
Metrics: success rate IK error rate
Baselines: ACT Diffusion Policy PerAct + Planner PerAct + Planner + Bezier PerAct + Diffuser RRT Pose Diffusion Joint Diffusion
Builds on: PerAct Diffusion Policy Diffuser ChainedDiffuser keyframe discovery
Closest papers by meaning
- Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation (CVPR 2025 · similarity 0.88)
- Hierarchical Equivariant Policy via Frame Transfer (ICML 2025 · similarity 0.87)
- SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution (CVPR 2024 · similarity 0.87)
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · similarity 0.86)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.86)
- AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile 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.