ACG: Action Coherence Guidance for Flow-based VLA models

TL;DR

ACG presents a training-free, test-time guidance strategy for flow-based VLA models to enhance action coherence. By constructing an incoherent action generation vector through identity attention and guiding sampling in the opposite direction, ACG improves intra-chunk coherence without retraining. Empirical results across RoboCasa, DexMimicGen, and SO-101 demonstrate substantial gains in success rates, particularly on fine-grained tasks, while maintaining action accuracy. The approach introduces a practical, plug-and-play mechanism to improve robustness in manipulation policies with minimal computational overhead and broad compatibility with existing flow-based models.

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.

Figures (7)

• Figure 1: Conceptual illustration of ACG. ACG constructs an incoherent vector field $v_\theta^\text{IC}$ and combines it with the original vector field $v_\theta$ to extrapolate a guidance vector that steers sampling toward coherent action sequences.
• Figure 2: Performance improvements of ACG over Vanilla GR00T-N1 bjorck2025gr00t, naïve multi-inference ensembling, and classifier-free guidance ho2022classifier on simulated manipulation benchmarks (RoboCasa robocasa, DexMG jiang2025dexmimicgen) and real-world pick-and-place tasks (SO-101 so101).
• Figure 3: Illustration of ACG. Based on the original inference procedure (left), we modify the self-attention layer by replacing the attention map with the identity map to generate incoherent action sequence (right). Finally, we guide the denoising vector with the opposite direction of the incoherent denoising vector.
• Figure 4: Visual demonstration of the simulation and real-world benchmarks used in our experiments. RoboCasa robocasa contains 24 atomic manipulation tasks in kitchen environments, grouped into seven skill domains. DexMG jiang2025dexmimicgen provides three bimanual embodiments for dexterous manipulation tasks. SO-101 so101 features real-world pick-and-place tasks, including strawberries and tic-tac-toe.
• Figure 5: Performance across the seven manipulation skill domains in RoboCasa robocasa. ACG shows strong improvements on fine-grained manipulation tasks including insertion and button pressing.