ACG: Action Coherence Guidance for Flow-based VLA models (2510.22201v1)

Abstract: Diffusion and flow matching models have emerged as powerful robot policies, enabling Vision-Language-Action (VLA) models to generalize across diverse scenes and instructions. Yet, when trained via imitation learning, their high generative capacity makes them sensitive to noise in human demonstrations: jerks, pauses, and jitter which reduce action coherence. Reduced action coherence causes instability and trajectory drift during deployment, failures that are catastrophic in fine-grained manipulation where precision is crucial. In this paper, we present Action Coherence Guidance (ACG) for VLA models, a training-free test-time guidance algorithm that improves action coherence and thereby yields performance gains. Evaluated on RoboCasa, DexMimicGen, and real-world SO-101 tasks, ACG consistently improves action coherence and boosts success rates across diverse manipulation tasks. Code and project page are available at https://github.com/DAVIAN-Robotics/ACG and https://DAVIAN-Robotics.github.io/ACG , respectively.

• The paper introduces Action Coherence Guidance (ACG), a novel test-time algorithm that steers flow-based VLA models away from incoherent action sequences.
• It perturbs self-attention layers to create an incoherent vector field and uses a tunable guidance scale to enhance temporal stability during inference.
• Experimental results demonstrate significant gains in fine-grained robotic tasks, with success rate improvements up to +28.8 percentage points.

Action Coherence Guidance (ACG) for Flow-based Vision-Language-Action Models

Introduction

The paper "ACG: Action Coherence Guidance for Flow-based VLA models" (2510.22201) introduces Action Coherence Guidance (ACG), a test-time guidance algorithm designed to enhance the temporal coherence of action sequences generated by flow-based Vision-Language-Action (VLA) models. These models, which leverage diffusion and flow matching techniques, have demonstrated strong generalization capabilities across diverse manipulation tasks. However, their high generative capacity makes them susceptible to noise in human demonstrations, resulting in temporally incoherent actions that degrade manipulation performance, especially in fine-grained tasks requiring precision.

ACG addresses this challenge by constructing an incoherent vector field through architectural perturbation and guiding the model away from incoherent behaviors during inference, thereby improving action stability and success rates without additional training.

Figure 1: Conceptual illustration of ACG, which constructs an incoherent vector field $v_\theta^\text{IC}$ and combines it with the original vector field to steer sampling toward coherent action sequences.

Methodology

Flow Matching Policy and Action Chunking

Flow matching policies model the conditional distribution of action chunks given multimodal inputs (visual observations, language instructions, and robot states). The policy is trained to denoise noisy action chunks via a learned vector field, using forward Euler integration during inference to generate temporally extended action sequences.

Action chunking, a common technique, generates multiple actions per inference step, reducing compounding errors and promoting smoother trajectories. However, intra-chunk incoherence remains a significant issue, as unstable motions within a chunk can lead to critical failures.

Action Coherence Guidance (ACG)

ACG is inspired by classifier-free guidance (CFG) but diverges in its approach by constructing an incoherent variant of the base model. Specifically, ACG replaces the attention map in selected self-attention layers with an identity matrix, severing temporal communication between action tokens. This architectural perturbation yields an incoherent denoising vector field $v_\theta^\text{IC}$.

During inference, ACG computes both the original and incoherent vector fields and extrapolates a guidance vector in the opposite direction of incoherence:

$$v_\theta^{\text{ACG}(\lambda)} = (1 + \lambda) v_\theta - \lambda v_\theta^\text{IC}$$

where $\lambda$ is the guidance scale. The resulting action chunk is generated by integrating this guided vector field, steering the policy toward temporally consistent actions.

Figure 2: ACG modifies the self-attention layer by replacing the attention map with the identity map to generate incoherent action sequences, then guides denoising in the opposite direction.

Implementation Details

ACG is implemented as a test-time algorithm and does not require retraining or access to additional data. The method is compatible with any transformer-based flow matching policy that utilizes self-attention over action tokens. In practice, only a subset of attention layers (typically the middle or later layers) are perturbed to balance computational cost and effectiveness. The guidance scale $\lambda$ and the number/position of perturbed layers are tunable hyperparameters.

Experimental Results

Benchmarks and Baselines

ACG was evaluated on RoboCasa (24 simulated kitchen tasks), DexMimicGen (bimanual dexterous manipulation), and real-world SO-101 pick-and-place tasks. Baselines included vanilla flow matching policies, action smoothing (Gaussian filtering), multi-inference ensembling, classifier-free guidance, and white noise guidance.

Figure 3: Performance improvements of ACG over Vanilla GR00T-N1, ensembling, and classifier-free guidance on simulated and real-world manipulation benchmarks.

ACG consistently outperformed all baselines, achieving substantial gains in success rates, particularly in fine-grained tasks such as button pressing (+23.1 pp), insertion (+11.8 pp), and real-world pick-and-place (+28.8 pp). Smoothing and ensembling methods yielded only modest improvements, often at the expense of action accuracy.

Figure 4: Visual demonstration of the simulation and real-world benchmarks used in the experiments.

Figure 5: Performance across the seven manipulation skill domains in RoboCasa, with ACG showing strong improvements in fine-grained manipulation.

Action Coherence Analysis

Quantitative metrics (Action Total Variation and JerkRMS) demonstrated that ACG produces the most temporally coherent actions among all methods, without sacrificing accuracy. Ensemble and smoothing methods improved coherence but often blurred task-relevant details, leading to suboptimal performance.

Figure 6: Qualitative comparison on grasping strawberries, showing ACG generates stable and accurate gripper trajectories, unlike baselines which exhibit temporal incoherence.

Ablation and Generalization

Hyperparameter analysis revealed that increasing the guidance scale and perturbing middle/later attention layers improved performance, but excessive perturbation degraded results. ACG was shown to be robust across different flow-based VLA models ($\pi_0$, SmolVLA), demonstrating its generalization capability.

Figure 7: (Left) Hyperparameter analysis of ACG; (Right) Comparison with Self-GAD, showing intra-chunk coherence (ACG) is more critical than inter-chunk coherence (Self-GAD), and combining both yields further improvements.

Practical and Theoretical Implications

ACG provides a practical, training-free solution for enhancing action coherence in flow-based VLA models, directly addressing the instability induced by noisy human demonstrations. The method is plug-and-play, requiring only architectural access at inference, and is compatible with a wide range of transformer-based policies.

From a theoretical perspective, ACG demonstrates that test-time guidance via architectural perturbation can effectively steer generative models away from undesirable behaviors, extending the utility of guidance strategies beyond image and video generation to robotic control. The results suggest that intra-chunk coherence is a key determinant of manipulation success, and that guidance strategies targeting temporal structure can yield significant gains.

Future Directions

Potential future research includes:

• Exploring more efficient guidance computation, such as perturbing only a minimal subset of layers or leveraging shared intermediate features to reduce overhead.
• Investigating other forms of architectural perturbation for guidance, including dropout, masking, or alternative attention mechanisms.
• Integrating ACG with inter-chunk coherence methods (e.g., Self-GAD) for further performance improvements.
• Extending guidance strategies to address other limitations of flow matching policies, such as robustness to domain shift or multi-agent coordination.

Conclusion

ACG introduces a principled, test-time guidance mechanism for improving temporal coherence in flow-based VLA models, yielding substantial gains in manipulation performance across simulation and real-world tasks. The approach is broadly applicable, computationally tractable, and highlights the importance of intra-chunk coherence for reliable robotic control. ACG opens new avenues for leveraging guidance strategies in generative policy models, with implications for both practical deployment and theoretical understanding of action generation in robotics.