Source-linked AI summary
Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies
Xiangdong Feng, Yuxuan Cheng, Chen Shi, Boyao Han, Yuxuan Yan, Yitong Hong, Zhuotao Tian, Li Jiang
TL;DR
Fixed chunk horizons leave open how long a predicted future should be trusted before replanning, even though predictable motion and precision-sensitive phases differ. DVAC uses final-step denoising variance and a rolling local-scale threshold to adapt execution; evaluations report higher success with fewer replans across diverse settings.
Problem
Fixed action-chunk execution lengths assume a constant trustworthy horizon across tasks, states, and phases, although execution length can substantially affect task success.
Source: §1 · p. 1Method
DVAC measures variance among final clean-action estimates, executes the stable low-variance prefix, and calibrates its threshold with a rolling local-variance estimate.
Source: §1 · p. 2Results
Takeaways & Limitations
Denoising stability is a simple inference-time signal for confidence-aware deployment of chunked robot policies.
Source: §5 · p. 8DVAC requires intermediate denoising trajectories and is not directly applicable to non-flow-based or non-diffusion-style policies; real-world evaluation remains limited in task diversity and embodiment scale.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Action chunking has become a common inference strategy for flow-based robot policies, improving action coherence by modeling multi-step temporal dependencies in demonstrations. However, the execution horizon is still typically set as an empirical fixed value, overlooking that predictable free-space motions and precision-critical interaction phases often require different replanning frequencies. In this work, we first show that the denoising process of flow-based policies contains an intrinsic signal of task phases: clean-action estimates remain stable during predictable motion phases, but fluctuate more strongly around contact-rich or precision-sensitive operations. Motivated by this observation, we propose DVAC (Denoising-Variance Adaptive Chunking), a test-time method that adaptively determines how many actions to execute from each predicted chunk. DVAC measures the variance of clean-action estimates over the final denoising steps, executes the stable low-variance prefix, and replans before high-variance future actions are committed. To transfer across tasks and rollouts, DVAC further calibrates the threshold with a rolling estimate of the local variance scale. Experiments on LIBERO, RoboTwin, CALVIN, and real-world manipulation show that DVAC improves task success while reducing replanning frequency. With a $\pi_{0.5}$-based policy, DVAC improves LIBERO success from 94.75% to 98.00% and reduces replanning by 43.0%, while also yielding aggregate gains on RoboTwin and CALVIN and improving real-world execution efficiency.
1 Introduction
Fixed action-chunk horizons leave policies with a trade-off between responsiveness and inference efficiency, despite task phases requiring different execution lengths. DVAC uses denoising variance to adapt replanning, with reported gains on LIBERO, RoboTwin, CALVIN, and real-world manipulation.
1 Introduction
- Fixed-horizon chunking assumes a constant trustworthy future across tasks, states, and phases, creating a trade-off between responsiveness and inference efficiency.The selected execution length can substantially affect task success.Source: §1 · p. 1
- Denoising variance stays low during moving phases and rises around contact-rich operations; operation phases show higher mean variance across LIBERO tasks.The analysis reports this phase pattern across all four LIBERO suites.Source: Figure 1 · p. 2, §1 · p. 2
- Prior adaptive approaches motivate asking whether a policy’s own inference process can signal when to stop executing a chunk and replan.Existing approaches often require extra candidate chunks, selection heuristics, or task-specific training.Source: §1 · p. 2
- 94.75% to 98.00% LIBERO success accompanies a 43.0% reduction in replanning frequency with DVAC and a π0.5-based policy.The method also uses a rolling estimate of local variance to calibrate its threshold across tasks and rollouts.Source: §1 · p. 2
- DVAC is a training-free test-time method that executes a stable low-variance prefix and replans before committing higher-variance future actions.Its contributions include an online scale-adaptive threshold and validation across LIBERO, RoboTwin, CALVIN, multiple backbones, and real-world manipulation.Source: §1 · p. 2
2 Related Work
Prior work uses chunked actions to improve temporal consistency and amortize inference, but fixed execution horizons trade feedback frequency against open-loop efficiency. DVAC instead reuses denoising variance from a single inference to adapt execution.
2 Related Work
- Diffusion and flow-based policies expose intermediate denoising states and clean-action estimates, but existing deployments typically use only the final chunk.DVAC reuses the denoising trajectory to decide how many actions to execute.Source: §2 · p. 3
- Most chunked policies use fixed execution horizons, while prior adaptive approaches include guided sampling, asynchronous inference, attention-based horizon estimation, entropy adaptation, and uncertainty-aware confidence estimation.Action chunking appears in ACT, Diffusion Policy, and robot foundation models.Source: §2 · p. 3
- Unlike approaches that use additional sampling or selection mechanisms, DVAC adapts execution from denoising variance in a single inference.Its overview monitors the final denoising steps to form a denoising-variance sequence, with a rolling buffer used to estimate local variance scale.Source: §2 · p. 3, Figure 2 · p. 3
3 Method
DVAC adapts receding-horizon execution by using variation among final clean-action estimates to identify how much of a predicted chunk to execute. A rolling local-variance threshold calibrates the rule across rollouts.
3 Method
- DVAC is a training-free strategy that uses denoising variance to adaptively decide when a flow-based policy should replan.The method section introduces flow-based action chunking, the denoising-variance execution rule, and scale-adaptive thresholding.Source: §3 · p. 3
3.1 Preliminaries: Flow-Based Action Chunking
- A flow-based policy predicts a horizon-H action chunk from state and executes only a prefix before observing a new state and replanning.The policy integrates a learned velocity field from noise to action space; intermediate clean-action estimates reveal prediction stability along that trajectory.Source: §3.1 · pp. 3–4
3.2 Denoising-Variance Adaptive Chunking
- DVAC estimates per-index variance from the final denoising steps; agreement among clean-action estimates marks stable future actions, while persistent fluctuation marks unstable ones.It executes the low-variance prefix before the first high-variance index, subject to minimum and maximum execution lengths.Source: §3.2 · p. 4
- A fixed variance cutoff can be too conservative or permissive across rollouts, so DVAC sets its threshold relative to a rolling estimate of the local variance distribution.The threshold is the rolling mean plus α times the rolling standard deviation; α sets tolerance in local-standard-deviation units.Source: §3.2 · pp. 4–5
- Under a local Lipschitz assumption, larger tail variance is linked to a larger denoising integration-error bound, motivating earlier replanning.The paper presents this as a rationale for using tail variance as a test-time signal.Source: §3.2 · p. 5
4 Experiments
Experiments assess task performance, backbone transfer, threshold sensitivity, and phase-adaptive execution across simulation benchmarks and real-world manipulation. DVAC improves benchmark outcomes and adapts execution horizons using denoising variance, with threshold calibration addressing differences in variance scale.
4.1 Simulation Experiments
- 0.980 average success: DVAC improves the π0.5 fixed-prefix LIBERO baseline from 0.948 and achieves the best average success rate.Table 1 compares four evaluation suites and their average, including prior adaptive chunking methods.Source: §4.1 · p. 5, Table 1 · p. 6
- The inference procedure estimates action-wise variance from the final L clean predictions, sets a scale-adaptive threshold, selects an executable prefix, and updates its rolling variance window.The algorithm returns the selected prefix of the final predicted action chunk.Source: §4.1 · p. 5
- RoboTwin average success rises from 0.359 to 0.416, while CALVIN-5 average completed subtasks increase from 3.905 to 4.040 with DVAC.Fixed thresholds reduce CALVIN-5 average completed subtasks to at most 3.716, while DVAC calibrates from recent variance statistics.Source: §4.1 · p. 6, Table 2 · p. 6, Table 3 · p. 6
- Average replanning falls from 32.6 to 18.6 for π0.5, from 32.8 to 23.0 for Qwen2.5-VL-π, and from 31.4 to 25.1 for Qwen3-VL-GR00T.The reported reductions accompany improved or preserved success, yielding a better efficiency–robustness trade-off than fixed-prefix execution.Source: §4.1 · p. 6, Table 4 · p. 7
4.2 Real-World Experiments
- DVAC improves real-world success over Fixed-15 while reducing task time and replanning; Fixed-40 is faster with fewer replans but substantially lowers success.Across the three real-world tasks, DVAC uses longer chunks during stable motion and shortens execution near contact-rich or precision-sensitive phases.Source: §4.2 · pp. 6–7, Table 5 · p. 7, Figure 4 · p. 7
4.3 Ablations
- LIBERO success remains stable across a broad range of fixed thresholds, suggesting the denoising-variance signal is not dependent on a fragile numerical value.Figure 5 examines different fixed threshold values across policy backbones.Source: §4.3 · p. 7, Figure 5 · p. 8
- Fixed-threshold transfer is unreliable across suites and backbones; CALVIN-5 AvgSub falls from 3.905 to at most 3.716, while DVAC reaches 4.040.The figure comparison also reports that DVAC has the best RoboTwin success rate and adaptive α yields more stable LIBERO phase correlations than fixed τ.Source: Figure 3 · p. 5, §4.3 · pp. 7–8, Figure 6 · p. 8
- Across 40 LIBERO task rollouts, adaptive thresholds show stable negative correlations around r < −0.27 with p < 0.05, indicating shorter chunks during operation-sensitive phases.Fixed-threshold behavior varies more with the absolute threshold.Source: §4.3 · p. 8, Figure 6 · p. 8
5 Conclusion
DVAC uses the variance of clean-action estimates over the denoising tail to choose how many actions to execute before replanning. Across diverse benchmarks, backbones, and real-world manipulation, it improves task success while reducing replanning overhead.
5 Conclusion
- DVAC measures tail variance of clean-action estimates to decide how many actions to execute from a predicted chunk before replanning.This turns the policy’s denoising trajectory into a test-time mechanism for adaptive execution.Source: §5 · p. 8
- Experiments span LIBERO, RoboTwin, CALVIN, multiple flow-based backbones, and real-world manipulation tasks.Source: §5 · p. 8
- Across diverse settings, DVAC improves task success while consistently reducing replanning overhead.The paper describes denoising stability as a simple signal for confidence-aware deployment of chunked robot policies.Source: §5 · p. 8
6 Limitations
DVAC treats denoising variance as an empirical proxy for action stability, not as calibrated uncertainty or a safety estimate. Its use requires intermediate denoising trajectories, and real-world evaluation remains limited in task diversity and embodiment scale.
6 Limitations
- Denoising variance is an empirical proxy for action stability, not a calibrated uncertainty or safety estimate.Source: §6 · p. 9
- DVAC assumes access to intermediate denoising trajectories and is not directly applicable to non-flow-based or non-diffusion-style policies.Source: §6 · p. 9
- Real-world evaluation remains limited in task diversity and embodiment scale.The authors plan further evaluation in more diverse and complex real-world manipulation settings.Source: §6 · p. 9
7 Appendix
The appendix develops a local stability rationale for variance-based execution and reports the evaluation setup, phase correlations, failure cases, and full evaluation results.
7.1 Proof of the Tail-Variance Error Bound
- The proof connects clean-action fluctuation to time-scaled finite differences of the learned vector field and states that tail variation is controlled by clean-action variance.The endpoint-error bound is monotone in variance under local assumptions, and is described as a stability argument rather than an absolute safety guarantee.Source: §7.1 · pp. 13–14
- The endpoint bound assumes local Lipschitz behavior over the denoising tail, and the appendix characterizes the resulting stability bound as conservative rather than a calibrated safety guarantee.Its constant depends on tail length, step size, tail interval, and the local Lipschitz constant.Source: §7.1 · p. 14
7.2 Additional Experimental Details
- The appendix evaluates LIBERO across several backbones, RoboTwin’s 16 clean-setting tasks, CALVIN, and three real-robot tasks, with phase analysis using one π0.5 rollout per LIBERO task.The real-world evaluation uses 50 demonstrations and 30 trials per task; the phase-correlation analysis covers 40 LIBERO tasks.Source: §7.2 · p. 15
7.3 Phase-Correlation Analysis
- Across LIBERO suites, OPERATING phases have higher median variance, and all tested α values yield negative phase–execution correlations around r < -0.27 (p < 0.05).Episode-level mean variance varies substantially, motivating scale-adaptive rather than fixed numerical thresholds.Source: §7.3 · p. 16, Figure 7 · p. 16, Figure 9 · p. 17
7.4 Failure Case Analysis
- On LIBERO, DVAC succeeds in 21 episodes where the fixed-prefix baseline fails, while the baseline succeeds in 8 episodes where DVAC fails.The episode-level comparison suggests adaptive execution more often corrects fixed-horizon failures than introduces new ones.Source: §7.4 · p. 16, Table 6 · p. 17
- Denoising stability can misalign with task correctness: stable but suboptimal chunks or late variance increases near contact can delay replanning and accumulate pose error.The appendix suggests combining variance with visual change, contact cues, or value estimates to improve robustness.Source: §7.4 · p. 17
7.5 Complete Evaluation Results
- Across configurations, fixed numerical thresholds show scale dependence, while adaptive DVAC is more stable; it reduces replanning without the reliability loss of very long fixed chunks.RoboTwin tasks also prefer different thresholds, reflecting variation in task phases and action uncertainty.Source: §7.5 · pp. 17–18, Table 7 · p. 19, Table 8 · p. 20