cs.ROMar 12, 2026
SaveNoise-Space Attribution and Control of Chunk-Boundary Artifact
Organizations: Nanjing University
Abstract
Action chunking is widely used in generative visuomotor policies, yet the recurring execution discontinuities at chunk boundaries still lack a mechanistic explanation. This paper treats chunk-boundary artifact as an analyzable mechanism variable. We first show that successful and failed episodes separate stably on artifact metrics. We then show that, in stochastic action-chunked policies, fixing the observation context and changing only latent noise is sufficient to modulate artifact systematically. On the same Diffusion Policy checkpoint, comparisons among DDPM, zero-variance DDPM, and DDIM further show that this local controllability depends on whether the information path from initial noise to action output remains intact. Finally, from controlled interventions at fixed local execution states, we find that artifact changes can carry through to final outcome, and that the preferred direction can reverse even within the same task: some contexts achieve higher success under lower artifact, whereas others achieve higher success under higher artifact. In a representative high-artifact-favoring key context selected by held-out matched-continuation validation, success rate increases from 0.033 to 0.717. These results show that chunk-boundary artifact is not a mere execution-side by-product, but a variable in noise space that can be attributed, controlled, and mechanistically linked to task outcome.
Figures & tables
| Task | Control condition | Success mean | Failure mean | value | |
|---|---|---|---|---|---|
| goal 3 | All episodes | 0.102 | 0.341 | +0.238 | |
| Contact-free | 0.100 | 0.297 | +0.197 | ||
| Contact-free first 50 steps | 0.093 | 0.138 | +0.045 | 0.005 | |
| goal 0 | All episodes | 0.108 | 0.219 | +0.110 | |
| Contact-free | 0.108 | 0.185 | +0.077 | 0.002 | |
| Contact-free first 50 steps | 0.049 | 0.070 | +0.021 | 0.008 |
Table 1: Episode-level boundary–interior jerk contrast for successful and failed rollouts. All values are from 20,000-sample permutation tests.
Figure 1: Matched-horizon action-jerk time courses. Even after truncating successful and failed rollouts to a common horizon, failed trajectories still exhibit stronger jerk pulses around replanning boundaries. Vertical dashed lines mark boundaries every five steps.
Figure 2: Artifact variation when the observation context is fixed and only latent noise is changed. Shaded regions indicate mean standard deviation across noise samples under the same context, and red dots denote reference rollouts. Top: boundary transition jerk. Bottom: local boundary–interior jerk contrast.
Figure 3: Local controllability along noise directions and information-path conditions. The horizontal axis is in all four panels, and each vertical axis is centered at , so the comparison is about response shape rather than absolute scale. (a) OpenPI with a near-linear response; (b) OpenPI bending at large ; (c) DDPM collapse; (d) ACT remaining nearly flat.
Figure 4: Real-robot directional scan on an Aubo i5 platform. Under a fixed execution context, sweeping along an OpenPI noise direction produces a monotone response in the centered first-boundary gap. The result supports hardware-side local controllability of chunk-boundary artifact, without making an outcome-level success claim.
| Policy state | Representative metric | Steering signal | Interpretation |
|---|---|---|---|
| ACT | first-boundary gap | no stable signed separation | weak residual artifact |
| OpenPI (flow matching; main experiments use pi0.5-libero ) | local sweep | strong local directional structure | |
| Diffusion Policy + DDPM | same weights | pathway largely broken | |
| Diffusion Policy + DDPM (subsequent variance set to zero) | same weights | still | removing reinjection does not restore directional structure |
| Diffusion Policy + DDIM ( ) | same weights | pathway intact |
Table 2: Mechanism-level comparison across policy states. The point is not direct comparability of absolute values across tasks, but the pattern: stochastic policies exhibit large and controllable artifact, the deterministic reference exhibits only weak residual artifact, and steering effectiveness depends on whether the noise pathway is accessible.
Figure 5: Full-trajectory steering. The top row shows a task whose baseline success rate is already near saturation: the mechanism metric is controllable, but positive success gains are compressed. The bottom row shows a task with clear room for improvement: target-low raises success from to , while target-high lowers it to ; meanwhile the boundary-gap metric preserves the consistent ordering low baseline high.
| Context | Preferred direction | Pairs | Baseline | Steered | Uplift | |
|---|---|---|---|---|---|---|
| target-low | 180 | |||||
| target-high | 60 |
Table 3: Paired steering results for two fixed key contexts within task 8. Both come from the same top-12 candidate list under matched-continuation validation, but their optimal intervention directions are opposite. The first row aggregates matched pairs and the second aggregates matched pairs. The baseline in the second row is the conditional success rate under the frozen reference noise at that key context, not the task-level average success rate.
| Statistic | Value |
|---|---|
| Variance decomposition of raw gap | context main effect ; noise main effect ; context noise interaction |
| Residual decomposition after subtracting baseline gap | context main effect ; noise main effect ; context noise interaction |
| Fraction of contexts where the globally best low-artifact noise remains locally optimal | |
| Mean gap under the globally best low-artifact noise | |
| Mean gap under context-wise optimal noise selection | |
| Relative adaptive gain | reduction in mean gap |
Table 4: Quantitative results from the larger fixed-context extension on task 8. The analysis uses 192 contexts and a shared set of 16 candidate noises to quantify the context noise interaction and to show why the optimal steering target cannot be treated as a global constant across contexts.
| Noise / Context | c1 | c2 | c3 | c4 | c5 | c6 |
|---|---|---|---|---|---|---|
| z1 | 0.199 | 0.166 | 0.029 | 0.044 | 0.105 | 0.102 |
| z2 | 0.169 | 0.158 | 0.018 | 0.067 | 0.126 | 0.061 |
| z3 | 0.169 | 0.183 | 0.012 | 0.073 | 0.112 | 0.065 |
| z4 | 0.208 | 0.212 | 0.011 | 0.073 | 0.103 | 0.047 |
| z5 | 0.109 | 0.096 | 0.046 | 0.080 | 0.119 | 0.142 |
| z6 | 0.224 | 0.249 | 0.024 | 0.059 | 0.055 | 0.050 |
Table A1 : Top-left slice of the OpenPI context–noise boundary-gap matrix used in the fixed-context noise analysis. Columns correspond to contexts through , and rows correspond to noise samples through .
| Task | Condition | BTJ std | gap std |
|---|---|---|---|
| goal 3 | vary only | 0.029 | 0.030 |
| vary only | 0.035 | 0.040 | |
| vary both | 0.038 | 0.038 | |
| goal 0 | vary only | 0.022 | 0.020 |
| vary only | 0.024 | 0.025 | |
| vary both | 0.017 | 0.019 |
Table A2: First-boundary noise decomposition. We vary only current-chunk noise , only next-chunk noise , or both together, and report the standard deviation of the first-boundary artifact metrics. BTJ denotes first-boundary transition jerk. Each condition uses 2 contexts 5 samples; goal 0 is overall less noise-sensitive than goal 3 ( vs ), and at this sample size the goal 0 “vary both” std is not separable from the single-factor conditions.
Figure A1 : Aggregate summary of the full-trajectory steering experiments ( per group). Left: success rate. Right: episode-level boundary–interior jerk contrast. The aggregate mixes near-saturation baselines, settings with substantial room for improvement, settings close to a failure floor, and settings with still-limited sample size; accordingly, pooled success-rate separation is compressed, while the directional ordering of the mechanism metric is preserved.
| Component | Result |
|---|---|
| First-stage sensitivity predictor | A linear classifier using only the current context predicts whether a local context is intervention-worthy, i.e. whether offline fixed-context validation shows that some intervention branch outperforms baseline there; under leave-one-trial-out evaluation it reaches accuracy and AUROC , above the majority-class baseline accuracy of |
| After connecting the first-stage filter to the online controller | the number of interventions drops to roughly – of the original level, but after aggregation over a longer window success drops from the baseline to , indicating that once the filter fires, second-stage dispatch—when to intervene, which direction to choose, and how to allocate budget—remains the main bottleneck |
| Online controller | a local positive result appears on a shorter evaluation window with , but after aggregation over a longer window it becomes only ; the one-sided sign test against baseline gives |
| Fixed-context predictor with 16 shared candidate noises | single-model selection reduces held-out mean gap from to ; the validation-selected top-3 ensemble yields positive gain on splits, with mean absolute gap gain |
Table A3: Supplementary diagnostics for adaptive steering. These are not presented as positive method results, but as key diagnostics: whether first-stage sensitivity screening is learnable, whether the online controller exhibits local positive signal, and whether adaptive selection becomes feasible at the artifact level before it becomes stable at the outcome level.
| Setting | Held-out selected gap | Global fixed baseline gap | Absolute gain | Note |
|---|---|---|---|---|
| single model (split seed 1) | 0.0815 | 0.1068 | 0.0253 | RMSE , Pearson |
| single model (5-split aggregate) | 0.0566 | 0.0763 | 0.0197 | positive gain on 4/5 splits |
| top-3 ensemble (5-split aggregate) | 0.0517 | 0.0763 | 0.0246 | positive gain on 5/5 splits |
Table A4: Results for the fixed-context predictor with 16 shared candidate noises. Single model denotes the best single model chosen by validation within each split; top-3 ensemble denotes the ensemble of the three models ranked highest by validation performance.
More in this topic
Visuomotor Policy Learning
Visuomotor policy learning trains robots to turn camera observations, often with language and proprioception, into actions for manipulation and navigation.
See all 286 papers