Research map / Reinforcement learning / Visual and representation RL
Rethinking Bimanual Robotic Manipulation: Learning with Decoupled Interaction Framework
Jian-Jian Jiang, Xiao-Ming Wu, Yi-Xiang He, Ling-An Zeng, Yi-Lin Wei, Dandan Zhang, Wei-Shi Zheng
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 decoupled design (independent model per arm) (ablation) | integrated control baseline · RoboTwin ablation (Flow Matching, 7 tasks) | average success rate: 0.426 -> 0.624 | Improved |
| add selective scaling factors on top of decoupled design (ablation) | decoupled framework · RoboTwin ablation (Flow Matching, 7 tasks) | average success rate: 0.624 -> 0.754 | Improved |
| add selective bias vectors on top of decoupled design (ablation) | decoupled framework · RoboTwin ablation (Flow Matching, 7 tasks) | average success rate: 0.624 -> 0.747 | Improved |
| add both scaling factors and bias vectors (ablation) | decoupled framework · RoboTwin ablation (Flow Matching, 7 tasks) | average success rate: 0.624 -> 0.789 | Improved |
| replace MLP interaction with selective interaction (ablation) | decoupled framework · RoboTwin (7 tasks) | average success rate: 0.736 -> 0.789 | Improved |
| replace concatenation interaction with selective interaction (ablation) | decoupled framework · RoboTwin (7 tasks) | average success rate: 0.733 -> 0.789 | Improved |
| apply decoupled interaction to PFM | Point Flow Matching · RoboTwin integration into existing methods | average success rate: 0.426 -> 0.789 | Improved |
| apply decoupled interaction to DP3 (DDIM) | DP3 · RoboTwin integration into existing methods | average success rate: 0.554 -> 0.674 | Improved |
| enlarge DP3 action generator from 262.43M to 576.43M | DP3 · RoboTwin model size | average success rate: 0.554 -> 0.361 | Hurt |
| shrink our model from 536.07M to 42.95M parameters | Decoupled Interaction Framework · RoboTwin model size | average success rate: 0.789 -> 0.719 | No clear effect |
| increase demonstrations from 150 to 200 | Decoupled Interaction Framework · RoboTwin Blocks Stack | average success rate: 0.530 -> 0.650 | Improved |
| use decoupled interaction instead of DP3 in real world | DP3 · Real-world (4 tasks, 60 trials) | success count: 27/60 -> 45/60 | Improved |
Closest papers by meaning
- Bi-DexHands: Towards Human-Level Bimanual Dexterous Manipulation (TPAMI 2023 · similarity 0.91)
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · similarity 0.90)
- Learning Diverse Bimanual Dexterous Manipulation Skills from Human Demonstrations (AAAI 2026 · similarity 0.90)
- ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.90)
- HumanoidGen: Data Generation for Bimanual Dexterous Manipulation via LLM Reasoning (NeurIPS 2025 · similarity 0.87)
- OAKINK2: A Dataset of Bimanual Hands-Object Manipulation in Complex Task Completion (CVPR 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.
