# VLA action chunk switching no longer stutters VLA action chunk switching no longer stutters Full article: https://haiguangboy.com/posts/learning_native_continuation_for_action_chunking_flow_policies ## Core method Continuation shouldn't be a patch—it has to be a native part of the policy's own denoising dynamics ↳ The diagnostic experiment first makes this clear: doing one-time prefix guidance only at the start of denoising causes actions in the overlap region to gradually drift away from the reference values over multiple denoising steps—proving that guidance must be reapplied at every step, and doing it only once is not enough ↳ Legato's approach: define a continuity vector spanning the action chunk's time course (strong guidance at the chunk start, gradually decaying to zero), and use it to reshape the flow-matching velocity field itself, making "step-by-step guidance" an intrinsic part of the denoising dynamics rather than an inference-time constraint wrapped around the standard velocity field ↳ The schedule is characterized by three parameters—latency, execution stride, and decay length—which are randomized during training and fed to the policy as explicit conditions, so a single model can adapt to different latency fluctuations ## Key results · On five real-robot manipulation tasks, it comprehensively outperforms RTC: shorter completion times, better scores on multiple smoothness metrics, and higher task completion scores—the smoothness gains were not bought at the cost of success rate · It also comprehensively beats the contemporaneous "training-time RTC" (which likewise introduces constraints during training, but only adds a hard prefix constraint without changing the underlying dynamics) · The method is not tied to a backbone: it was validated on π0, π0.5, GR00T N1.6, and even a UNet-based simulation policy ## Comparison with related lines of work - Opposite line · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: Core line-of-work judgment: inter-chunk continuation should be a native property of the policy's own learned denoising dynamics, not an external mechanism patched on outside the standard velocity field via hard constraints at inference time or training time—this is the fundamental stance that distinguishes Legato from RTC and training-time RTC contradicts training-time RTC: masking the first d action tokens so the model learns smooth continuation - Same line · [PointWorld: point flow unified state-action](https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation) `pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08`: PointWorld: point flow unified state-action - Same line · [Which layer should RL intervene at](https://haiguangboy.com/posts/zprl) `beyond_action_residuals_real_world_robot_policy_steering_via_bottleneck_latent_r_2026_08`: Compared with the contemporaneous 'training-time RTC' (which also introduces continuation constraints during training, but merely adds a hard prefix constraint to the overlapping action segment without changing the underlying flow dynamics itself), Legato comprehensively wins on task score, completion time, and smoothness—showing that what really matters is reshaping the policy's own denoising dynamics, not simply adding a hard training-time constraint validates problem restatement: the key to RL post-training is not just 'how much to change,' but 'at which layer to intervene' - Same line · [Qianxun Intelligence Han Fengtao: Embodied intelligence enters the "racing for speed" stage](https://haiguangboy.com/posts/qianxun-general-brain-data-loop) `wx_China_Wealth_20260719_2026_07`: The paper's author completed this work during an internship at Qianxun Intelligence (Spirit AI) (team lead: Gao Yang), and it was accepted to the top robotics conference RSS 2026; Legato is named after the musical term "legato"—smooth, unbroken transitions between notes validates ★Team combination: industry veteran + Abbeel disciple, Gao Yang's lineage directly connected to the founders of Physical Intelligence/Covariant - Same line · [Blending RGB, 3D geometry, and object semantics into a single representation—this is the fundamental reason it is more accurate than π0.5 and Fast-WAM](https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility) `flex_pi_a_multi_stream_world_action_model_with_compute_flexibility_2026_08`: FLEX-π: a more comprehensive joint representation - Same line · [The more accurate the future prediction, the more sluggish the robot becomes](https://haiguangboy.com/posts/omega-0) `omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08`: Action chunking is standard for VLA deployment, amortizing inference cost and supporting high-frequency control, but inference latency plus the inherent multimodality of flow policies often makes the junction between adjacent action chunks unsmooth, manifesting as hesitation and abrupt turns, and prolonging task completion time validates deployment: a single forward pass takes about 0.14 seconds (over 7Hz), predicts a 25-step action chunk but executes only the first 8 steps, and RTC-style warm-starting with overlap blending smooths the continuation ## Boundaries · The NLDLJ metric, which measures high-order smoothness over the entire trajectory, did not improve significantly on all tasks—the improvements are concentrated near chunk boundaries, and do not mean the whole trajectory is smoother everywhere · Robustness to deployment latency comes with preconditions: for latency values not covered by the training-time randomization range (especially the extreme case of "zero latency"), performance may degrade severely or even approach complete failure · The validation scope is limited to continuous action chunking policies like flow-matching; VLA architectures with discrete action representations are not covered ## Author's judgment (not part of the paper; a cross-paper synthesis view) "Randomize during training the troublesome parameters that will vary during deployment, and let the policy adapt on its own"—this technique is homologous to the previously interpreted PointWorld: PointWorld randomizes the number of cameras, this paper randomizes inference latency, and the result in both cases is that the model becomes most robust to various values of that parameter. The trade-off between stride and decay length in the schedule ablation is also the same pattern as PAVE's non-monotonic finding on the number of prediction anchors: giving a system more constraints or parameters is not unconditionally good; the key is finding that just-right trade-off point.