SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models

AI summary

Overview

• Research area: Robotics — Vision-Language-Action (VLA) models, uncertainly-aware inference, and test-time scaling for closed-loop robotic manipulation.
• Technical level: Intermediate. Readers should be comfortable with autoregressive token decoding, sampling temperature, KL divergence, and the general VLA pipeline (vision encoder plus language-conditioned action decoder).
• Scope: The paper introduces SCALE, a training-free, single-forward-pass inference strategy that jointly modulates visual attention and action sampling in VLA models using a self-uncertainty signal derived only from output logits, and evaluates it on three simulated benchmarks and a real-world robot.

What This Paper Is About

VLA models are usually run with a fixed inference pipeline: a frozen vision encoder processes the observation, and greedy decoding commits to the single top-1 action token at each step. Existing test-time scaling (TTS) methods for VLAs improve robustness but require extra training, an external verifier, and multiple forward passes, and they adjust only action decoding while leaving visual representations untouched. This paper asks whether the model's own predictive uncertainty can be used, with no extra training and only one forward pass, to adapt both what to do (action sampling) and how to perceive (visual attention) at test time.

Key Contributions

1. A dual-reference self-uncertainty measure. The authors define two reference distributions — a one-hot distribution on the top-1 token (q^low, full certainty) and a uniform distribution over the vocabulary (q^high, full ambiguity) — and score uncertainty as u^k_t = D_KL(p^k_t || q^low) - D_KL(p^k_t || q^high). This is shown to equal the expected log-likelihood ratio between the two references under the model's own distribution, and it captures both overall distributional spread and decisiveness about the top-1 choice.

2. Uncertainty-conditioned adaptive action decoding. Per-token sampling temperature is set as τ^k_t = T_0 · σ(u^k_t), so decoding is near-greedy under low uncertainty and explorative under high uncertainty. The sigmoid is interpreted as the posterior probability of the "uncertain" hypothesis, acting as a soft gate.

3. Uncertainty-conditioned adaptive visual attention. Step-level uncertainty u_t is the average of token-level uncertainties; an exponential moving average ū_t tracks recent history; the deviation Δu_t = u_t - ū_{t-1} is converted to a vision-encoder attention temperature γ_t = κ^{tanh(Δu_{t-1})}, which is applied across all layers to scale self-attention. Positive deviation broadens attention, negative deviation sharpens it.

4. A joint, training-free, single-pass inference loop. The two mechanisms are coupled across timesteps (action decoding at step t produces the uncertainty used for visual modulation at step t+1), with no additional training, no verifier, and one forward pass per control step.

Main Findings

• Consistent gains over greedy decoding across backbones and benchmarks. On LIBERO, SCALE improves average success rate by +5.8 with OpenVLA (75.7 to 81.5) and +1.8 with π0-FAST (91.2 to 93.0). On SIMPLER-WidowX it adds +14.6 with π0-FAST (34.4 to 49.0), +3.4 with fine-tuned SpatialVLA (41.7 to 45.1), and +10.4 with zero-shot SpatialVLA (31.3 to 41.7). On LIBERO-PRO-Long it adds +3.5 with OpenVLA (18.0 to 21.5) and +3.1 with π0-FAST (35.7 to 38.8).

• Notable long-horizon improvement on LIBERO with OpenVLA. SCALE reaches 63.3 on the Long split versus 52.7 for greedy OpenVLA, 54.7 for temperature sampling, 55.2 for top-k (k=40, t=0.7), and 55.1 for top-p (p=0.9).

• Real-world gains are larger than simulated ones. On the UR10e "Put A on B" tasks, SCALE raises in-distribution average success from 36.1 to 55.6 (+19.5) with OpenVLA and from 72.2 to 86.1 (+13.9) with π0-FAST. Under out-of-distribution conditions it raises averages from 22.9 to 39.6 (+16.7) with OpenVLA and from 43.8 to 56.3 (+12.5) with π0-FAST.

• Naive decoding strategies often hurt. With π0-FAST on LIBERO, temperature sampling, top-k, and top-p all underperform greedy decoding (91.2 versus 84.3, 88.1, and 86.2). The authors attribute this to fixed hyperparameters that cannot adapt across states and tasks.

• SCALE beats training-required TTS methods while staying training-free and single-pass. On LIBERO with OpenVLA, SCALE averages 81.5 versus 70.8 for MG-Select (+10.7), with 63.3 versus 55.4 on long-horizon tasks. It also surpasses reported LIBERO-Long numbers for RoboMonkey (56.5) and TACO (60.0) by +6.8 and +3.3 respectively.

• The two components are synergistic, not merely additive. In the ablation on OpenVLA/LIBERO-Long, adaptive decoding alone yields 58.0 (+5.3) and adaptive visual attention alone yields 56.0 (+3.3) against the 52.7 greedy baseline, but combining them reaches 63.3 (+10.6), exceeding the summed individual gains of +8.6 by over 2%.

• The dual-reference measure outperforms other uncertainty proxies. Under the same pipeline: p_max 56.0, Self-Certainty 53.8, Entropy 55.4, Gini Impurity 57.8, and the proposed self-uncertainty 63.3 — a 5.5-point margin over the next best.

• Design choices for visual modulation matter. Modulating uni-modal attention in the vision encoder beats cross-modal attention in the VLA backbone (63.3 versus 57.4, +5.9); adaptive continuous scaling beats fixed sign-based switching (57.4 versus 54.8, +2.6); and the change-based signal Δu_{t-1} beats the instantaneous u_{t-1} (63.3 versus 55.4, +7.9).

• Qualitative evidence of mechanism. Attention visualizations from SigLIP (the OpenVLA vision encoder) at t=45 show the fixed-temperature baseline attending to task-irrelevant regions such as the microwave door, while SCALE increases γ_t and redirects attention to the target mug, enabling a successful grasp. In a real-world π0-FAST task, greedy decoding collides with a bowl, whereas SCALE's elevated exploration selects a trajectory that clears the obstacle.

Methodology in Plain English

The authors treat the VLA policy's own output distribution as the only signal they need. At each decoded action token, the model produces a probability distribution over the action vocabulary. Instead of asking how close that distribution is to uniform (as prior LLM work does), they ask how it sits between two extremes: a perfectly peaked distribution on the top choice, and a perfectly flat distribution. Measuring the gap between the two KL divergences gives a single number that rises when the model is unsure overall or unsure about its top pick.

That number drives two knobs. For action decoding, it sets the sampling temperature per token: near-zero temperature (essentially greedy) when the model is confident, and up toward a maximum T_0 when it is not. For perception, the model averages token uncertainties into a per-step score, keeps a smoothed running average, and looks at whether the current step's uncertainty is above or below that running average. If it is above, the vision encoder's self-attention is flattened (γ_t > 1) so the model takes in more of the scene; if below, attention is sharpened (γ_t < 1). Because the step's uncertainty is only known after action decoding, the visual modulation for step t uses the previous step's deviation, which the authors justify by noting consecutive visual frames are highly correlated and report that using the current step's value with an extra forward pass yields similar performance (Appendix D).

No component is trained. Hyperparameters are tuned once on LIBERO-Long per backbone and then reused unchanged across all other simulated benchmarks and the real robot. Exact values of T_0, α, κ, and ε are not reported in the paper content provided; the paper points to Appendix L for hyperparameter sensitivity analysis.

Why This Matters

• Research impact. The work reframes test-time scaling for VLAs: instead of spending compute on extra rollouts and verifiers, it spends nothing extra and modulates the inference process itself. It also extends uncertainty-based decoding beyond LLM-style distributional confidence to a measure that also captures top-1 decisiveness, which matters when an action is executed irreversibly. Its empirical claim that adaptive scaling beats fixed-temperature sampling challenges a common default in VLA evaluation.

• Potential real-world applications:

  • Pick-and-place manipulation in logistics or warehouse settings, where distractors and novel object geometry are common.
  • Household or service robots handling soft, irregular, or unseen objects (the OOD tasks in this paper include a soft teddy bear and a small cube).
  • Industrial or lab automation with a fixed arm, since SCALE is evaluated on a 6-DoF UR10e with a Robotiq 2F-85 gripper.
  • Deployment on latency-constrained robots, because it keeps a single forward pass per control step rather than multiple samples.

• Industry relevance. VLA deployment is often limited by real-time constraints and by the cost of training or maintaining verifier networks. A training-free method that reuses logits already computed during decoding avoids both, and it applies on top of existing fine-tuned checkpoints (OpenVLA, π0-FAST, SpatialVLA) rather than requiring new architectures.

Future Directions

• Hyperparameter transfer. The paper tunes SCALE's hyperparameters on LIBERO-Long only and reuses them elsewhere; whether a single setting transfers to entirely different embodiments or action spaces is left open.
• Extension beyond the vision encoder. The authors choose uni-modal attention modulation in the vision encoder over cross-modal attention in the VLA backbone, and report a +5.9 gap. Whether a hybrid or hierarchical scheme could do better is untested.
• Temporal proxy. Visual modulation uses the previous step's uncertainty deviation because the current step's is unavailable before encoding. The paper reports that using the current value with an extra forward pass gives similar performance, but it does not explore learned or multi-step predictors of scene complexity.
• Scope of the uncertainty signal. SCALE is evaluated on autoregressive VLAs with discretized action tokens; applicability to non-autoregressive or continuous-action policies is not investigated in the content presented.

Target Audience

Robotics and embodied-AI researchers working on VLA models and manipulation policies, particularly those interested in inference-time methods that avoid retraining. It is also relevant to practitioners deploying VLAs on real hardware under latency constraints, and to readers interested in uncertainty quantification for generative sequence models, since the dual-reference log-likelihood-ratio formulation is presented as a general alternative to uniform-divergence confidence measures. Readers without background in autoregressive decoding or KL divergence will need to consult the cited priors on sampling and self-certainty first.

Authors’ abstract

Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic control, with test-time scaling (TTS) gaining attention to enhance robustness beyond training. However, existing TTS methods for VLAs require additional training, verifiers, and multiple forward passes, making them impractical for deployment. Moreover, they intervene only at action decoding while keeping visual representations fixed-insufficient under perceptual ambiguity, where reconsidering how to perceive is as important as deciding what to do. To address these limitations, we propose SCALE, a simple inference strategy that jointly modulates visual perception and action based on 'self-uncertainty', inspired by uncertainty-driven exploration in Active Inference theory-requiring no additional training, no verifier, and only a single forward pass. SCALE broadens exploration in both perception and action under high uncertainty, while focusing on exploitation when confident-enabling adaptive execution across varying conditions. Experiments on simulated and real-world benchmarks demonstrate that SCALE improves state-of-the-art VLAs and outperforms existing TTS methods while maintaining single-pass efficiency.