Denoising Tells When to Replan: Denoising-Variance Adaptive Chunking for Flow-Based Robot Policies

Technical Summary: Denoising-Variance Adaptive Chunking (DVAC)

1. Problem Statement

Flow-based robot policies, such as those utilizing Diffusion or Flow Matching, have become standard for generating continuous action sequences. A common deployment strategy is action chunking, where the policy predicts a sequence of future actions, and a prefix of this sequence is executed before the policy replans. While chunking improves temporal coherence and amortizes inference costs, current implementations rely on a fixed execution horizon (e.g., always executing the first 10 actions of a predicted 50).

This fixed approach creates an inherent trade-off:

• Predictable phases (e.g., free-space arm movement) often require less frequent replanning to maintain efficiency.
• Precision-critical phases (e.g., contact-rich interactions, object manipulation) require frequent replanning to correct errors.
  A static horizon cannot adapt to these varying dynamics, leading to either inefficient over-replanning or risky under-replanning during critical tasks. Existing adaptive methods often require additional candidate chunks, extra selection heuristics, or task-specific training, complicating deployment.

2. Methodology: DVAC

The authors propose Denoising-Variance Adaptive Chunking (DVAC), a training-free, test-time framework that dynamically determines the execution horizon based on the internal dynamics of the flow-based policy's inference process.

Core Insight:
During flow matching inference, the policy iteratively refines a noisy action chunk into a clean action estimate through a denoising trajectory. The authors observe that the variance of these intermediate clean-action estimates ($\hat{x}_0$) across the final denoising steps serves as an intrinsic signal of task phase stability:

• Low Variance: Occurs during predictable motion phases, indicating stable predictions.
• High Variance: Occurs during contact-rich or precision-sensitive operations, indicating uncertainty and the need for replanning.

Algorithm Overview:

1. Variance Estimation: For a predicted action chunk of horizon $H$, DVAC monitors the final $L$ steps of the denoising process. It calculates the variance ($V_s(k)$) of the clean-action estimates for each future action index $k$.
2. Thresholding: Instead of using a fixed numerical threshold (which varies across tasks and rollouts), DVAC employs a scale-adaptive threshold ($\tau_s$). It maintains a rolling buffer of recent variance statistics to compute a local mean ($\mu_s$) and standard deviation ($\sigma_s$). The threshold is set as $\tau_s = \mu_s + \alpha\sigma_s$, where $\alpha$ is a hyperparameter defining uncertainty tolerance.
3. Adaptive Execution: The policy executes the longest prefix of actions where the variance remains below the threshold. Once the variance exceeds $\tau_s$ (indicating an upcoming unstable phase), execution stops, and the policy triggers a replan.
4. Constraints: The execution length is bounded by minimum ($N_{min}$) and maximum ($N_{max}$) limits to ensure stability and prevent excessive replanning.

3. Key Contributions

1. Identification of Intrinsic Signal: The paper establishes that denoising convergence stability is an intrinsic, training-free signal for adaptive execution in flow-based policies, directly correlating with contact-rich and precision-sensitive task phases.
2. DVAC Framework: The proposal of a plug-and-play, training-free inference-time method that adapts execution horizons using denoising variance and an online scale-adaptive threshold.
3. Empirical Validation: Comprehensive evaluation across multiple benchmarks (LIBERO, RoboTwin, CALVIN), diverse flow-based backbones (including $\pi_0.5$, Qwen2.5-VL, and GR00T), and real-world manipulation tasks.

4. Experimental Results

The authors evaluated DVAC on simulation benchmarks and real-world hardware, comparing it against fixed-horizon baselines and other adaptive methods.

• LIBERO Benchmark: Using a $\pi_0.5$-based policy, DVAC improved the average success rate from 94.75% to 98.00% while reducing the replanning frequency by 43.0% (from 32.6 to 18.6 replans per episode). It outperformed other adaptive baselines like AAC and AutoHorizon.
• RoboTwin & CALVIN: DVAC improved the average success rate on RoboTwin from 0.359 to 0.416 and increased the average completed subtasks on CALVIN-5 from 3.905 to 4.040. Notably, fixed-threshold variants underperformed on CALVIN, highlighting the necessity of DVAC's scale-adaptive design.
• Backbone Transferability: The method demonstrated "plug-and-play" capabilities, improving performance across different model architectures (e.g., Qwen2.5-VL and Qwen3-VL-GR00T) without retraining.
• Real-World Manipulation: In tasks involving placing cubes, stacking, and moving test tubes, DVAC consistently achieved higher success rates than fixed-horizon baselines (e.g., 0.867 vs. 0.800 for cube placement) while significantly reducing task time and replanning counts.
• Phase Correlation: Analysis confirmed that DVAC successfully shortens execution horizons during operation-sensitive phases and extends them during moving phases, aligning with the theoretical motivation.

5. Significance and Claims

The paper claims that DVAC offers a simple yet effective mechanism for confidence-aware deployment of chunked robot policies. By leveraging the denoising trajectory already computed during inference, DVAC eliminates the need for:

• Additional training.
• Extra candidate action generation.
• Task-specific tuning of numerical thresholds.

The authors argue that denoising stability is a readily available, intrinsic signal that allows policies to balance dynamic responsiveness (replanning when necessary) with inference efficiency (executing longer chunks when stable). This approach improves task success while simultaneously reducing the computational overhead of frequent replanning, addressing a critical bottleneck in the deployment of large-scale visuomotor policies.

Limitations Acknowledged:
The authors note that DVAC uses denoising variance as an empirical proxy rather than a calibrated uncertainty estimate. It is currently limited to flow-based or diffusion-style policies that expose intermediate trajectories and assumes access to these internal states. Real-world evaluation was limited in task diversity and embodiment scale.