Research map / Graph learning / Knowledge and temporal graphs
Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual 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 KStar Diffuser, a diffusion policy for bimanual manipulation that conditions denoising on a spatial-temporal robot-structure graph (GCN over URDF-derived joint graph) and a differentiable forward-kinematics regularizer, outperforming ACT, PerAct2, DP and DP3 on RLBench2 and a real ALOHA setup.
Problem
Bimanual imitation learning policies that predict next-best end-effector poses and then apply inverse kinematics often produce self-collisions, inter-arm interference, and joint-limit-violating poses.
Approach
Model the dual-arm robot's physical structure as a dynamic spatial-temporal graph built from URDF joints and feed it as a condition to the diffusion denoiser, and add a differentiable forward-kinematics regularizer that predicts joint angles and maps them to reference end-effector poses to condition the diffusion process on kinematic feasibility.
Key results
- On RLBench2 with 100 demos, KStar Diffuser reaches 68.2% overall success versus 40.5% for DP-EE and 33.9% for PerAct2.
- In real-world ALOHA experiments, KStar Diffuser achieves 43.1% overall success versus 29.9% for PerAct2.
- Ablation on handover-item shows overall success dropping from 23.4% (full) to 16.8% without kinematics regularizer and 14.8% without both components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the spatial-temporal robot graph and kinematics regularizer | KStar Diffuser · RLBench2 overall, 20 demos | success rate: 24.9 ±2.1 -> 58.0 ±1.4 | Helped |
| add the spatial-temporal robot graph and kinematics regularizer | KStar Diffuser · RLBench2 overall, 100 demos | success rate: 33.9 ±1.0 -> 68.2 ±2.1 | Helped |
| add the spatial-temporal robot graph and kinematics regularizer | KStar Diffuser · Real-world overall | success rate: 29.9 ±10.2 -> 43.1 ±17.8 | Helped |
| add the kinematics regularizer (ablation) | KStar Diffuser with ST Graph · Handover-S | success rate: 18.3 ±1.5 -> 27.0 ±1.7 | Helped |
| add the kinematics regularizer (ablation) | KStar Diffuser with ST Graph · Handover-R | success rate: 15.3 ±3.3 -> 19.7 ±5.3 | Helped |
| add the kinematics regularizer (ablation) | KStar Diffuser with ST Graph · Overall | success rate: 16.8 ±2.1 -> 23.4 ±5.2 | Helped |
| add the spatial-temporal graph and kinematics regularizer (ablation) | KStar Diffuser without ST Graph and KR · Handover-S | success rate: 16.3 ±3.1 -> 27.0 ±1.7 | Helped |
| add the spatial-temporal graph and kinematics regularizer (ablation) | KStar Diffuser without ST Graph and KR · Handover-R | success rate: 13.3 ±9.4 -> 19.7 ±5.3 | Helped |
| add the spatial-temporal graph and kinematics regularizer (ablation) | KStar Diffuser without ST Graph and KR · Overall | success rate: 14.8 ±2.1 -> 23.4 ±5.2 | Helped |
Surprising findings
- Joint-space diffusion baseline DP-J performs worst among diffusion methods (21.5% overall at 100 demos), below end-effector-based DP-EE (40.5%).
Limitations
- Real-world evaluation is limited to two tasks with 15 trials each
- Kinematics regularizer and graph require URDF-specific structure and joint-level supervision
Experimental setup
Datasets: RLBench2
Benchmarks: RLBench2 real-world ALOHA tasks (Lift Plate, Handover)
Metrics: success rate
Baselines: ACT RVT-LF PerAct-LF PerAct2 DP-J DP-EE DP3
Builds on: Diffusion Policy RLBench2 PerAct2 ACT
Closest papers by meaning
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · similarity 0.90)
- Hierarchical Diffusion Policy for Kinematics-Aware Multi-Task Robotic Manipulation (CVPR 2024 · similarity 0.88)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.87)
- GraphMimic: Graph-to-Graphs Generative Modeling from Videos for Policy Learning (CVPR 2025 · similarity 0.87)
- ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY (ICLR 2025 · similarity 0.86)
- VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 2024 · 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.