PaperlayerView exact revision on arXiv

Source-linked AI summary

Mixture of Horizons in Action Chunking

Dong Jing, Gang Wang, Jiaqi Liu, Weiliang Tang, Zelong Sun, Yunchao Yao, Zhenyu Wei, Yunhui Liu, Zhiwu Lu, Mingyu Ding

arXiv:2511.19433v2cs.ROcs.AIcs.CV

TL;DR

VLA policies are highly sensitive to the action horizon, with short horizons favoring precision and long horizons favoring foresight, leaving fixed choices suboptimal. Mixture of Horizons combines multiple horizons through shared processing and gated fusion, while enabling consensus-based dynamic inference. Across simulation and real-world evaluations, it consistently improves policies and reaches 99% average LIBERO success for π0.5 after 30k mixed-task iterations.

  • Problem

    VLA performance is highly sensitive to action horizon, while fixed horizons impose a trade-off between long-term foresight and short-term precision.

  • Method

    Mixture of Horizons segments action chunks across multiple horizons, processes them with a shared action transformer, fuses predictions through a lightweight gate, and supports consensus-based dynamic inference.

  • Results

    Across simulation environments and real-world tasks, MoH consistently improves flow-based and one-step regression policies; π0.5 with MoH achieves 99% average success on LIBERO after 30k mixed-task iterations.

  • Takeaways & Limitations

    MoH provides a plug-and-play strategy that combines long-term foresight with short-term precision and supports more stable, efficient execution.

  • Takeaways & Limitations

    Many remaining LIBERO failures are attributed to environmental issues or instruction-following limitations outside MoH’s design scope.

Abstract

from arXiv · show

Vision-language-action (VLA) models have shown remarkable capabilities in robotic manipulation, but their performance is sensitive to the $\textbf{action chunk length}$ used during training, termed $\textbf{horizon}$. Our empirical study reveals an inherent trade-off: longer horizons provide stronger global foresight but degrade fine-grained accuracy, while shorter ones sharpen local control yet struggle on long-term tasks, implying fixed choice of single horizons being suboptimal. To mitigate the trade-off, we propose a $\textbf{mixture of horizons (MoH)}$ strategy. MoH rearranges the action chunk into several segments with different horizons, processes them in parallel with a shared action transformer, and fuses outputs with a light linear gate. It has three appealing benefits. 1) MoH exploits long-term foresight and short-term precision jointly within a single model, improving both performance and generalizability to complex tasks. 2) MoH is plug-and-play for full-attention action modules with minimal training or inference overhead. 3) MoH enables dynamic inference with adaptive horizons, which selects stable actions through cross-horizon consensus, achieving 2.5$\times$ higher throughput than baselines while preserving superior performance. Extensive experiments over flow-based policies $π_0$, $π_{0.5}$, and one-step regression policy $π_{\text{reg}}$ demonstrate that MoH yields consistent and significant gains on both simulations and real-world tasks. Notably, under mixed-task setting, $π_{0.5}$ with MoH reaches a new state-of-the-art with 99$\%$ average success rate on LIBERO after only $30k$ training iterations. Project page: https://timsty1.github.io/moh/

1. Introduction

VLA performance depends strongly on action horizon: shorter horizons improve precise short-term control, whereas longer horizons provide foresight for long-term tasks. Mixture of Horizons combines multiple horizons to alleviate this trade-off and improves results across policies, tasks, and inference settings.

  • 1. Introduction: Longer horizons improve long-term planning, while shorter horizons provide more precise control, making a fixed single horizon an inherent generalization bottleneck.The study evaluates horizons in [10, 20, 30] on LIBERO task suites ranging from short to long trajectories.
  • 1. Introduction: MoH segments each action chunk into different horizons, processes them in parallel with a shared action transformer, and fuses predictions using a lightweight linear gate.The gating layer adds only 2k parameters, and horizons [10, 20, 30] jointly mitigate the observed trade-off.
  • 1. Introduction: MoH is plug-and-play for full-attention action modules and introduces minimal training and inference overhead.The strategy applies to both flow-based and single-step prediction-based action modules.
  • 1. Introduction: Cross-horizon consensus enables dynamic inference by selecting consistently supported actions, forming self-truncating chunks and deferring uncertain actions.The mechanism improves execution stability and efficiency; π0.5 with MoH surpasses baseline π0.5 even at 2.5× throughput.
  • 1. Introduction: 99% average success rate on LIBERO is achieved by π0.5 with MoH after 30k iterations in mixed-task training, establishing a new state of the art.The paper reports consistent improvements across flow-based and single-step policies, simulation environments, and real-world robotic tasks.

2. Related Work

Prior work uses action chunking to predict future action sequences, but horizon selection is typically heuristic and performance varies by task type. Existing methods do not provide a general way to resolve the resulting precision–foresight trade-off.

  • 2. Related Work: Action chunking predicts future action sequences instead of single next actions, supporting temporal structure, high-frequency control, and smoother execution.Overlapping actions at matching timestamps can be fused, and CogACT refines this with similarity-based weighting.
  • 2. Related Work: Horizon selection is typically heuristic, despite evidence that performance is highly sensitive to horizon and different task types favor different horizons.This sensitivity motivates a universal mixture-of-horizons training strategy.
  • 2. Related Work: Prior work lacks an available method for mitigating the trade-off between long-term foresight and short-term precision induced by a fixed horizon.The paper addresses this gap with a universal mixture-of-horizons training strategy.

3. Method

MoH combines multiple action-chunk horizons within a shared full-attention action transformer, using gated fusion to balance long-term foresight and short-term precision. It also supports consensus-based dynamic inference and training objectives for both flow-based and one-step policies.

  • 3.2. Mixture of Horizons: A fixed horizon trades long-term planning against precise short-term control, motivating the use of multiple horizons within one policy.Longer horizons perform better on long-horizon tasks, while shorter horizons favor short-horizon tasks.
  • 3.2. Mixture of Horizons: MoH processes all horizon variants in parallel after padding them to the maximum length and applying horizon-specific attention masks.The VLM prefix is computed once, while invalid positions beyond each horizon are masked.
  • 3.2. Mixture of Horizons: MoH rearranges each action chunk into segments with different horizons, processes them in parallel using a shared action transformer, and fuses valid predictions with a linear gate.The gate produces per-step, per-horizon weights and adds only 2k parameters.
  • 3.2. Mixture of Horizons: The shared action transformer produces horizon-specific predictions, while masked gate logits are normalized over horizons valid at each step before fusion.This leaves the backbone unchanged and applies to both flow-matching and one-step policies.
  • 3.3. Dynamic Inference via Horizon Consensus: Cross-horizon consensus forms a self-truncating executable prefix by committing actions supported across horizons and deferring uncertain actions to replanning.The dynamic scheme is designed to improve execution stability and inference-time efficiency.

4. Simulation Experiments

MoH improves VLA performance across simulation benchmarks and ablations by combining diverse horizons, while dynamic inference improves efficiency without sacrificing performance.

  • LIBERO: 99% average success rate makes π0.5 with MoH a new LIBERO state of the art after 30k iterations.MoH consistently improves all four LIBERO suites.
  • RoboTwin: π0 with MoH achieves the highest average success rate on RoboTwin and improves over base π0 on most tasks.Gains appear on both easy and hard variants despite training only on easy demonstrations.
  • RoboCasa: GR00T with MoH improves on all five RoboCasa tasks, yielding a 3.4% average gain across a less-saturated environment.The result also transfers across distinct VLA backbones.
  • Ablation Study: 99.0% success is reached with stride d = 3, while three horizons already raise average success from 97.7% to 98.3%.Denser horizons generally help, but more horizon groups are not always better.
  • Ablation Study: 99.0% with diverse horizons outperforms 97.9% from ten identical H=30 branches, showing gains primarily come from horizon diversity.Identical branches also fail to alleviate the Long-suite trade-off.
  • Dynamic Inference: 2.5× throughput still surpasses baseline π0.5 under dynamic inference, which executes cautiously or quickly according to action uncertainty.Throughput measures executed action steps per predicted chunk rather than wall-clock latency.

5. Real-world Experiments

Real-robot experiments evaluate MoH across short-horizon, long-horizon, and fine-grained manipulation tasks. MoH consistently improves performance and combines more decisive short-term control with more accurate long-horizon behavior.

  • Result and Analysis: MoH yields consistent gains across all four real-world tasks and both π0 and π0.5, improving long-horizon decisions and short-horizon precision.The evaluation covers bread placement, milk pouring, drawer manipulation, and diagonal towel folding.
  • Result and Analysis: π0.5 with MoH improves towel folding from 75% (15/20) to 90% (18/20), benefiting phase-sensitive precise control.
  • Result and Analysis: On long-horizon drawer manipulation, MoH produces more accurate grasps, quicker completion, and higher success rates.
  • Result and Analysis: On bread placement, MoH replaces baseline hesitation with a faster, more direct grasping motion.
  • Result and Analysis: MoH alleviates π0.5’s milk-pouring hesitation between continuing to pour and returning the bottle, which the authors associate with local-conflict overfitting.
  • Result and Analysis: The real-world results are consistent with simulation findings, supporting MoH’s combination of long-horizon planning and short-term control.

6. Conclusion

The paper introduces MoH as a plug-and-play multi-horizon strategy for full-attention VLA policies. Across simulation and real-world evaluations, it consistently improves flow-matching and one-step regression policies and reaches a new LIBERO state of the art.

  • MoH consistently improves flow-matching and one-step regression policies across simulator benchmarks and real-world tasks, achieving a new state of the art on LIBERO.
  • MoH fuses multi-horizon action chunks in full-attention VLA policies to ease the trade-off between long-term foresight and short-term precision.
  • Ablations support dense horizons, gated fusion, gating-balance regularization, and dynamic inference as beneficial components.

Impact Statement

The paper identifies single-horizon action chunking as a critical VLA limitation because it trades off long-term foresight against short-term control fidelity. MoH is presented as an efficient, broadly applicable approach that alleviates this trade-off and supports adaptive closed-loop control.

  • A single action-chunk horizon induces a trade-off between long-term foresight and short-term control fidelity in VLAs.
  • MoH is plug-and-play and computationally efficient, consistently alleviating the horizon trade-off while enabling flexible dynamic inference for adaptive closed-loop control.
  • The authors argue that MoH’s simple design generalizes broadly across base models and diverse settings.
  • The approach is suggested as a potential pathway toward more robust VLA pretraining and scalable foundational action models.

A. Training hyperparameters

This appendix lists training hyperparameters for the π series on LIBERO and reports timing costs under different horizon settings. MoH adds little training or inference overhead, with inference latency virtually unchanged.

  • Table 6 reports the training hyperparameters used for π0, π0.5, and πreg on LIBERO.
  • MoH brings very little additional training and inference time for π0 and π0.5 under different horizon settings.
  • Inference latency is virtually unaffected by MoH, preserving the control frequency and usability of VLA models.

C. Discussion of MoH Hyper-parameter Setup in Pre-training and Post-training

MoH supports sparse horizons for efficient pre-training and denser horizons for performance-focused post-training, while Hmax=30 offers the strongest LIBERO results among tested settings.

  • Pre-training: A sparse three-horizon configuration adds approximately 1.4% training overhead while retaining foresight and precision benefits.
  • Post-training: A denser horizon configuration increases training overhead by roughly 10% but improves fine-grained control and long-term planning stability for complex manipulations.
  • Dynamic inference: Dynamic inference executes shorter chunks near decision points and fine-grained manipulation, reserving longer chunks for smooth, low-risk motion.
  • Horizon trade-off: π0.5 shows the same horizon trade-off as π0: longer horizons improve long-horizon performance, while MoH substantially improves overall success across tasks.
  • Hmax selection: Hmax=30 achieves the superior LIBERO results among MoH configurations with 10 horizon groups.Smaller Hmax favors fine-grained control but lacks global foresight, whereas excessively large values can weaken local precision.

F.3. Statistical Analysis of Gating Balance Loss

The gating-balance analysis links longer training horizons to lower short-term motor precision and shows that Lbal prevents systematic preference for longer horizons.

  • Gating balance: Without Lbal, gating weights favor longer horizons because those chunks participate in more steps, creating statistical and gradient bias.
  • Gating balance: Lbal suppresses this bias and enables more balanced horizon utilization during gating.
  • Trade-off: Longer horizons improve foresight but sacrifice short-term motor precision, establishing the per-step form of the action-chunking trade-off.
  • Motor precision: Per-step ℓ1 error increases monotonically with training horizon on Spatial, Object, and Goal suites, directly indicating reduced low-level precision.

G.2. Task-Adaptive Gating Weights

MoH learns task-adaptive horizon allocation: longer-horizon weights rise as task complexity increases, while disagreement-based thresholds adapt dynamic execution to changing phases.

  • Task-adaptive weights: As LIBERO complexity increases from Spatial to Long, weight shifts from h=3 toward h=30, matching the horizon trade-off.The shortest-horizon weight changes from 9.09 to 8.93, while the longest changes from 10.21 to 10.24.
  • Dynamic thresholds: The dynamic scheme computes a threshold from the first n chunk steps, adapting to the current task and policy state without additional learning.
  • Dynamic thresholds: Cross-horizon disagreement ranges from 0.056 in aligned segments to 0.110 near challenging transitions, motivating data-dependent truncation.
  • Task-adaptive weights: Shorter horizons receive greater support near difficult transitions, while smoother phases permit more consistent longer-horizon execution.

I.1. LIBERO

LIBERO and RoboTwin evaluations reveal both strong qualitative behavior and important evaluation boundaries, including simulator mislabeling, occlusion, perception errors, and low-level precision failures.

  • LIBERO: π0.5 with MoH would reach 99.8% on LIBERO-Spatial after excluding two environment-induced errors, versus 98.8% reported in the main table.
  • Failure modes: Remaining failures include target misidentification, insufficient low-level precision, out-of-view arm states, and illegible randomized object views.
  • Scope and extensions: MoH does not directly address perception and language-understanding errors, while memory and refined dynamic gating remain proposed extensions.
  • RoboTwin limitations: RoboTwin can fail to signal completion after successful behavior because success conditions are overly strict and position thresholds are miscalibrated.
  • RoboTwin limitations: Severe occlusion can hide the target object, making state inference difficult and motivating richer history information.
  • Qualitative demonstrations: MoH produces precise low-level motions and completes long-horizon, multi-stage goals in qualitative LIBERO, real-world, and RoboTwin demonstrations.
Loading 2511.19433v2…