Research map / Vision-language and multimodal / Robot action and navigation
DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter
Xukun Li, Yu Sun, Lei Zhang, Bo-Sheng Huang, Yibo Peng, Yuan Meng, Haojun Jiang, Shaoxuan Xie, Guocai Yao, Alois Knoll, Zhenshan Bing, Xinlong Wang and 1 more
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 |
|---|---|---|---|
| decoupled DiT conditioning (joint self-attention for vision, AdaLN for proprioception) instead of Diffusion Policy FiLM conditioning | DECO vs DP · DECO-50 all tasks (vision only) | success rate (%): 51.25% -> 72.25% | Improved |
| add the plugin tactile adapter to pretrained vision-based DECO | DECO · DECO-50 all tasks | success rate (%): 72.25% -> 82.50% | Improved |
| add the tactile adapter on contact-rich tasks | DECO · Waste Disposal and Assembly | success rate (%): 53.13% -> 73.13% | Improved |
| add tactile via adapter on weakly tactile-relevant tasks | DECO · Pick and Place | successes out of 120: 103/120 -> 108/120 | No clear effect |
| concatenate tactile into the input of Diffusion Policy trained from scratch | DP · Assembly | successes out of 80: 15/80 -> 15/80 | No clear effect |
| concatenate tactile into DP input | DP · Waste Disposal | successes out of 80: 30/80 -> 38/80 | Improved |
| couple tactile with proprioception conditioning (add tactile embeddings to AdaLN) instead of decoupled cross-attention (ablation) | DECO from scratch · Assembly ablation stage 3 | successes out of 40: 29/40 -> 21/40 | Hurt |
| add tactile via decoupled cross-attention trained from scratch (ablation) | DECO · Assembly ablation stage 3 | successes out of 40: 18/40 -> 29/40 | Improved |
| use plugin adapter (7.97M trainable params) instead of training from scratch (89.41M) (ablation) | DECO tactile model · Assembly ablation stage 3 | successes out of 40: 29/40 -> 29/40 | No clear effect |
| swap pathways: proprioception via cross-attention and tactile via AdaLN (ablation) | DECO.ds · Assembly ablation stage 3 | successes out of 40: 29/40 -> 0/40 | Hurt |
| swap pathways: proprioception via cross-attention and tactile via AdaLN (ablation) | DECO.ds · Assembly ablation stage 2 | successes out of 40: 34/40 -> 17/40 | Hurt |
| add tactile (adapter or concatenation) on small smooth objects in sorting | DECO and DP · Material Sorting, object 00081-p-2x | successes out of 20: DECO 10/20 -> DECO.p 13/20; DP 5/20 -> DP.t 12/20 | Improved |
Closest papers by meaning
- RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation (ICLR 2025 · similarity 0.87)
- Enhancing Tactile-based Reinforcement Learning for Robotic Control (NeurIPS 2025 · similarity 0.87)
- AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile Manipulation (NeurIPS 2025 · similarity 0.87)
- Touch in the Wild: Learning Fine-Grained Manipulation with a Portable Visuo-Tactile Gripper (NeurIPS 2025 · similarity 0.86)
- DIFFTACTILE: A Physics-based Differentiable Tactile Simulator for Contact-rich Robotic Manipulation (ICLR 2024 · similarity 0.86)
- ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 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.
