Research map / Image generation / Fast diffusion sampling
STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction
Jinhao Li, Yuxuan Cong, Yingqiao Wang, Hao Xia, Shan Huang, Yijia Zhang, Ningyi Xu, Guohao Dai
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 velocity-aware perturbation injection (sigma_stall 0 -> 1.2) to STEP (ablation) | STEP, 2 steps · Real-world PickNPlace | success rate: 0.00 -> 0.95 | Improved |
| add velocity-aware perturbation injection (sigma_stall 0 -> 1.2) to STEP (ablation) | STEP, 2 steps · Real-world StackCube | success rate: 0.00 -> 0.80 | Improved |
| increase perturbation scale sigma_stall to 2.0 (ablation) | STEP · Real-world PickNPlace and StackCube | success rate: 1.00 -> 0.00 (step=8, sigma_stall 1.2 -> 2.0) | Hurt |
| use moderate predictor weight sigma_scale 0.2 or 0.4 instead of 0 (ablation) | STEP, 2 steps, sigma_stall=1.2 · Real-world PickNPlace | success rate: 0.65 -> 0.95 | Improved |
| increase predictor weight sigma_scale from 0.2 to 1.0 (ablation) | STEP, 2 steps, sigma_stall=1.2 · Real-world PickNPlace | success rate: 0.95 -> 0.65 | Hurt |
| use STEP warm-start instead of DDIM at 2 steps | Diffusion policy · Real-world PickNPlace | success rate: 0.60 -> 0.95 | Improved |
| use STEP warm-start instead of DDIM at 4 steps | Diffusion policy · Real-world StackCube | success rate: 0.80 -> 1.00 | Improved |
| use STEP instead of DDIM at 2 steps (state-based) | Diffusion policy · Transport (state-based) | score: 0.04 -> 0.86 | Improved |
| use STEP instead of BRIDGER at 2 steps (state-based) | Diffusion policy · ToolHang (state-based) | score: 0.08 -> 0.64 | Improved |
| use STEP instead of BRIDGER at 2 steps (ManiSkill2 PushChair) | Diffusion policy · ManiSkill2 PushChair | score: 0.45 -> 0.39 | Hurt |
| use STEP instead of BRIDGER at 4 steps (image-based Transport) | Diffusion policy · Transport (image-based) | score: 0.88 -> 0.88 | No clear effect |
| reduce denoising steps from 4 to 2 in STEP (state-based Push-T) | STEP · Push-T (state-based) | score: 0.91 -> 0.49 | Hurt |
Closest papers by meaning
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.92)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.92)
- Falcon: Fast Visuomotor Policies via Partial Denoising (ICML 2025 · similarity 0.91)
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.90)
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.89)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · 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.
