Stochastic Diversity Windows: Diagnosing Noise Influence in Robotic Diffusion Sampling
Organizations: School of Computer Science and Informatics, Cardiff University, UK · School of Engineering, Cardiff University, UK
Abstract
Diffusion models are increasingly adopted in robotics and control due to their capability to express multiple feasible solutions to a given task. However, the role of stochasticity in the diffusion model-sampler across denoising remains poorly understood. We introduce a per-timestep counterfactual diagnostic that tests how changing the sampled noise affects the final output. Across five model-sampler configurations spanning three robotics benchmarks, diverse architectures, domains, and conditioning mechanisms, we reveal highly varied temporal sensitivity profiles, where noise-induced deviations can amplify, partially contract, or undergo near-total attenuation downstream. From these profiles, we define stochastic diversity windows (SDWs): compact intervals that identify where sampler stochasticity has the strongest measured influence on the final output. Guided by these windows, we apply two heuristic interventions: targeted stochastic amplification, which increases sample diversity exploration of the solution space, and an SDW-hybrid sampler, which replaces DDPM steps with DDIM skips in low-impact regions, resulting in network evaluation reductions of up to 40-60% and wall-clock time by up to 85%, with no consistent changes detected in feasibility, cost, or diversity across held-out tasks.
Figures & tables
| Maze2D | MPD-Prior | MPD-Guided | PushT- Transf. | PushT-CNN | |
| Domain | 2D Nav. | Motion Plan | Motion Plan | visuomotor policy | visuomotor policy |
| Sampling object | trajectory | motion plan | motion plan | target-pos. | target-pos. |
| Diffused shape | |||||
| Architecture | U-Net | U-Net | U-Net | Transf. | FiLM U-Net |
| Training loss | |||||
| Prediction target |
| Maze2D | MPD | PushT | |||
|---|---|---|---|---|---|
| Prior | Guided | Transf. | CNN | ||
| Task conditions | 10 | 10 | 5 | 10 | 10 |
| Baseline samples/task | 10 | 10 | 10 | 10 | 10 |
| Timesteps probed | 114 | 25 | 25 | 100 | 100 |
| Repeats/timestep | 10 | 50 | 10 | 10 | 10 |
| Batch size | 50 | 50 | 10 | 1 | 1 |
| Model | Amplification ( , ) | Feasibility | Diversity | Mean cost | Best cost |
|---|---|---|---|---|---|
| Maze2D | Baseline ( ) | ||||
| MPD-P | Baseline ( ) | ||||
| Feasible-sample diversity change | ||||
|---|---|---|---|---|
| Model | Diagnosed (Ours) | Middle | Opposite | |
| MPD-G | ||||
| MPD-P | ||||
| PushT-T | ||||
| Sampler | NFEs | Wall-clock (s) | Feasibility/50 | Best Cost | Diversity |
|---|---|---|---|---|---|
| MPD-P (inpainting only) | |||||
| Full DDPM | 25 | ||||
| SDW-hybrid, | 10 | (-85%) | (+5%) | (0%) | (+10%) |
| DDIM, | 10 | (-85%) | (-38%) | (+32%) | (+182%) |
| SDW-hybrid, | 14 | (-79%) | (+4%) | (+2%) | (+2%) |
| DDIM, | 14 | (-79%) | (-21%) | (+16%) | (+150%) |
| Setting | Placement | NFEs (DDPM/DDIM) | Feasibility/50 | Cost | Diversity |
|---|---|---|---|---|---|
| MPD-P, | SDW | 10 (5/5) | |||
| MPD-P, | Middle | 10 (5/5) | |||
| MPD-P, | SDW | 14 (10/4) | |||
| MPD-P, | Middle | 14 (10/4) | |||
| MPD-G, | SDW | 15 (12/3) | |||
| MPD-G, | Middle | 15 (12/3) |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| 0.50 | 47 | 31.5 | 34.22 | 44 | 32.0 | 33.95 | 48 | 29.5 | 34.33 | 47 | 28.8 | 34.53 |
| 0.75 | 46 | 33.2 | 34.22 | 45 | 33.0 | 34.25 | 46 | 30.0 | 34.16 | 48 | 30.5 | 34.34 |
| 1.00 | 44 | 35.2 | 34.19 | 44 | 35.2 | 34.19 | 44 | 35.2 | 34.19 | 44 | 35.2 | 34.19 |
| 1.50 | 45 | 37.0 | 34.15 | 38 | 39.4 | 33.24 | 38 | 43.8 | 33.34 | 39 | 46.8 | 33.96 |
| 2.00 | 40 | 40.8 | 33.85 | 37 | 44.9 | 32.91 | 35 | 47.7 | 32.91 | 33 | 47.8 | 33.55 |
| MPD-P | ||||
|---|---|---|---|---|
| Baseline ( ) | ||||
| PushT Transformer | |||
|---|---|---|---|
| Baseline ( ) | |||
| Configuration | SDW-hybrid | Middle control | NFEs |
|---|---|---|---|
| MPD-P, | – | – | 10 |
| MPD-P, | – | – | 14 |
| MPD-G, | – , – | – , – | 15 |
Diffusion Sampling
Diffusion sampling studies how to turn a trained diffusion model into high-quality samples efficiently and reliably.
See all 291 papers