Research map / Image generation / Fast diffusion sampling
On-Device Diffusion Transformer Policy for Efficient Robot Manipulation
Yiming Wu, Huan Wang, Zhenghao Chen, Jianxin Pang, Dong Xu
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 learnable layer pruning (ablation) | MDT-V · CALVIN D->D | latency (ms): 22.25 -> 18.87 | Improved |
| add learnable layer pruning (ablation) | MDT-V · CALVIN D->D | average rollout length: 3.72 -> 3.70 | No clear effect |
| add consistency distillation (ablation) | MDT-V · CALVIN D->D | latency (ms): 22.25 -> 11.34 | Improved |
| add consistency distillation (ablation) | MDT-V · CALVIN D->D | average rollout length: 3.72 -> 3.69 | No clear effect |
| combine pruning and distillation (ablation) | MDT-V · CALVIN D->D | latency (ms): 22.25 -> 8.70 | Improved |
| prune DP-T to 2 layers with 4 steps | DP-T · Push-T | success rate: 0.754 -> 0.724 | No clear effect |
| prune DP-T to 2 layers with 4 steps | DP-T · Push-T | inference speed (ms): 90.6 -> 0.97 | Improved |
| reduce local block size M from 4 to 2 at depth 4 (ablation) | DP-T · Push-T | success rate: 0.747 -> 0.732 | Hurt |
| prune MDT-V deeper (E1-D1) | MDT-V · CALVIN ABCD->D | average rollout length: 4.52 -> 3.44 | Hurt |
| prune MDT-V to E3-D3 | MDT-V · CALVIN ABCD->D | average rollout length: 4.52 -> 4.50 | No clear effect |
| prune MDT-V to E3-D3 | MDT-V · LIBERO (average) | success rate (%): 74.3 -> 73.2 | No clear effect |
| prune DP-T to D6/6-8 | DP-T · real-world Task 1 | success rate: 0.80 -> 0.75 | Hurt |
Closest papers by meaning
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.90)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.90)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.89)
- AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile Manipulation (NeurIPS 2025 · similarity 0.89)
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.88)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.88)
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.
