Reliable uncertainty estimates are important for adaptive execution in vision-language-action (VLA) models. Existing approaches such as Adaptive Action Chunking (AAC) estimate action entropy from multiple sampled action chunks, introducing additional inference cost and capturing both uncertainty within an action trajectory and diversity across alternative strategies. We introduce Divergence-based Uncertainty-aware Adaptive Action Chunking (DUAC), a post-hoc method that extracts trajectory-conditioned predictive uncertainty from the vector field of a pretrained VLA. Building on a closed-form flow-posterior covariance formulation, DUAC derives action-level uncertainty scores from a single generated chunk using Jacobian-vector products, without retraining or multiple endpoint samples. DUAC strongly correlates with AAC, while capturing distinct information: AAC measures global dispersion across sampled chunks, whereas DUAC measures local uncertainty around the trajectory selected for execution. Consequently, DUAC is less sensitive to diversity among distinct but valid action strategies. Across LIBERO, LIBERO-Pro, LIBERO-Plus, and RoboCasa, DUAC-based adaptive horizon selection performs favorably among the evaluated methods, combining strong task success with low uncertainty-estimation latency. These results establish the flow vector field as an efficient, trajectory-conditioned source of predictive uncertainty for robot decision making.
Figures & tables
Fig. 1: Uncertainty over task progress for successful and failed LIBERO-10 rollouts. Curves show mean ±1 standard deviation. Frame rows show representative outcomes.
Fig. 2: Method overview of divergence-based uncertainty estimation for flow-based action generation and its use in adaptive execution-horizon selection.
TABLE II: Performance and computational efficiency comparison on LIBERO-Pro.
Method
Reloc.
Contact
Rot.
Button
Overall
Mean H
Dec./ep.
Fixed-25
58.80%
65.33%
26.80%
43.00%
51.79%
25.00
18.98
Fixed-50
54.90%
65.83%
30.80%
32.33%
49.79%
50.00
9.67
AutoHorizon
51.60%
68.83%
27.80%
37.67%
49.21%
34.96
13.84
Entropy AAC
51.20%
59.83%
25.80%
43.67%
47.12%
26.80
18.43
DUAC (ours)
57.20%
66.67%
31.00%
43.00%
52.33%
29.58
16.17
TABLE III: RoboCasa Kitchen-24 results: success rate (%), mean execution horizon (Mean H), and mean policy decisions per episode (Dec./ep.).
Fig. 3: Multiple valid strategies increase marginal entropy more strongly than trajectory-conditioned divergence. The illustrations contrast alternative sampled trajectories with uncertainty around one selected trajectory; the right panel reports normalized uncertainty over the first three chunks.
Fig. 4: Mean policy decisions over task progress.
Fig. 5: Divergence uncertainty under increasing LIBERO-Pro position perturbations. Curves show mean uncertainty at different flow times, error bars show 95% confidence intervals, and success rates are shown below the perturbation levels.
Fig. 6: Pearson correlation between entropy and divergence across task progress and per-action-step uncertainty profiles. Boxes show the mean ±1 standard deviation.
Method
SR
Decisions
Mean H
Entropy AAC
75%
58.01
7.53
DUAC
75%
49.48
8.83
TABLE IV: Execution statistics on the two-moka-pot task over 100 rollouts.
Task
SR–AAC
SR–DUAC
Decisions–AAC
Decisions–DUAC
Decisions vs. AAC
OpenSingleDoor
50%
52%
16.13
14.65
−1.48
OpenDoubleDoor
36%
62%
30.54
24.88
−5.66
CloseSingleDoor
73%
83%
13.08
11.43
−1.65
CloseDoubleDoor
35%
37%
19.84
17.02
−2.82
TABLE V: Success rates and mean policy decisions per episode on matched RoboCasa door tasks. Decision differences are computed as DUAC minus AAC.
Method
Spatial
Object
Goal
LIBERO-10
Average
Fixed-10
99.4%
99.6%
97.8%
92.8%
97.4%
AutoHorizon
98.1%
98.3%
97.8%
93.7%
96.98%
AAC
98.8%
99.8%
98.2%
93.0%
97.5%
DUAC (ours)
99.4%
99.2%
98.4%
94.8%
98.0%
TABLE VI: Success rates (%) with a π0.5 backbone on LIBERO.
Perturbation
Fixed-10
AutoHorizon
AAC
DUAC (ours)
Background
97.00%
96.75%
96.45%
96.50%
Camera
60.70%
60.75%
63.30%
63.85%
Language
85.20%
84.65%
84.95%
85.35%
Light
95.50%
95.70%
95.70%
96.15%
Objects layout
85.20%
85.15%
85.70%
85.75%
Robot init.
83.15%
82.85%
80.90%
81.70%
TABLE VII: Success rates (%) with a π0.5 backbone across LIBERO-Plus perturbation categories.
Action chunking has recently emerged as a standard practice in flow-based Vision-Language-Action (VLA) models. However, the effect and choice of the execution horizon - the number of actions to be executed from each predicted chunk - remains underexplored. In this work, we first show that varying the execution horizon leads to substantial performance deviations, with performance initially improving and then declining as the horizon increases. To uncover the reasons, we analyze the cross- and self-attention weights in flow-based VLAs and reveal two key phenomena: (i) intra-chunk actions attend invariantly to vision-language tokens, limiting adaptability to environmental changes; and (ii) the initial and terminal action tokens serve as stable anchors, forming latent centers around which intermediate actions are organized. Motivated by these insights, we interpret action self-attention weights as a proxy for the model's predictive limit and propose AutoHorizon, the first test-time method that dynamically estimates the execution horizon for each predicted action chunk to adapt to changing perceptual conditions. Across simulated and real-world robotic manipulation tasks, AutoHorizon is performant, incurs negligible computational overhead, and generalizes across diverse tasks and flow-based models.
Vision-language-action models (VLAs) combine vision-language backbones with expressive generative action heads trained via flow matching on large-scale robotic datasets. Despite their strong empirical performance in robotic manipulation, VLAs lack mechanisms to quantify confidence in their predictions and to detect when their actions may be unreliable. This presents a critical limitation for real-world deployment in non-stationary environments, where models inevitably encounter scenarios outside their pretraining distribution and may fail without warning. To address this, we derive an efficient method for quantifying epistemic uncertainty in flow-matching models by leveraging velocity-field disagreement (VFD) across a small ensemble. We successfully use this uncertainty estimate for failure detection during deployment and active fine-tuning of flow-based VLAs. To this end, we propose SAVE, a framework for uncertainty-guided active multitask fine-tuning that reduces the number of costly expert demonstrations required to adapt VLAs to new tasks. Through extensive experiments on the LIBERO benchmark, we demonstrate that VFD yields better-calibrated uncertainty estimates predictive of downstream performance, that VFD achieves strong performance in detecting failures, and that uncertainty-guided data acquisition with SAVE requires at least 22% fewer samples than baselines. In summary, our work shows that quantifying epistemic uncertainty in flow-based VLAs improves both failure awareness and adaptation. Project website: tum-lsy.github.io/uq_vla/.
Vision-Language-Action (VLA) models enable general-purpose robotic policies by mapping visual observations and language instructions to low-level actions, but they often lack reliable introspection. A common practice is to compute a token-level uncertainty signal and take its mean over a rollout. However, mean aggregation can dilute short-lived but safety-critical uncertainty spikes in continuous control. In particular, successful rollouts may contain localized high-entropy segments due to benign noise or non-critical micro-adjustments, while failure rollouts can appear low-entropy for most timesteps and only exhibit brief spikes near the onset of failure. We propose a unified uncertainty quantification approach for predicting rollout success versus failure that (1) uses max-based sliding window pooling to preserve transient risk signals, (2) applies motion-aware stability weighting to emphasize high-frequency action oscillations associated with unstable behaviors, and (3) performs DoF-adaptive calibration via Bayesian Optimization to prioritize kinematically critical axes. Experiments on the LIBERO benchmark show that our method substantially improves failure prediction accuracy and yields more reliable signals for failure detection, which can support downstream human-in-the-loop interventions.