Research map / Image generation / Autoregressive and latent generation
Pixel Motion Diffusion is What We Need for Robot Control
E-Ro Nguyen, Yichi Zhang, Kanchana Ranasinghe, Xiang Li, Michael S. Ryoo
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 Motion Director pixel motion to the Action Expert (ablation) | Enhanced DP · CALVIN ABC->D | Avg. Len: 2.78 -> 4.00 | Improved |
| use pixel motion instead of RGB goal image conditioning (ablation) | DAWN · CALVIN ABC->D | Avg. Len: 3.21 -> 4.00 | Improved |
| initialize Motion Director from pretrained latent diffusion model (ablation) | DAWN · CALVIN ABC->D | Avg. Len: 3.42 -> 4.00 | Improved |
| add gripper view conditioning to Motion Director (ablation) | DAWN · CALVIN ABC->D | Avg. Len: 3.74 -> 4.00 | Improved |
| add gripper view (ablation) | VPP · CALVIN ABC->D | Avg. Len: 3.58 vs 3.93 with gripper view | Improved |
| increase Motion Director diffusion steps from 2 to 25 (ablation) | DAWN · CALVIN ABC->D | Avg. Len: 3.88 -> 4.00 | Improved |
| increase diffusion steps from 25 to 40 (ablation) | DAWN Motion Director · CALVIN ABC->D | Avg. Len: 4.00 -> 3.95 | No clear effect |
| add external DROID robot data to training | DAWN · CALVIN ABC->D | Avg. Len: 4.00 -> 4.10 | Mixed |
| add pixel motion from Motion Director to the Action Expert (ablation) | Enhanced Diffusion Policy · Galaxea bimanual test set | MSE: 0.128 -> 0.117 | Improved |
| use two-stage design with Motion Director | DAWN · real-world inference time | inference time (ms): 112.77 (Enhanced DP) -> 319.82 | Hurt |
Closest papers by meaning
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.91)
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.90)
- Conditioning Matters: Training Diffusion Policies is Faster Than You Think (NeurIPS 2025 · similarity 0.90)
- Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.90)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.89)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
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.
