Source-linked AI summary
VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model
Wenhao Li, Xiu Su, Yichao Cao, Hongyan Xu, Xiaobo Xia, Shan You, Yi Chen, Chang Xu
TL;DR
VLA models often act quickly without deliberation, which can cause poor or catastrophic decisions in complex or ambiguous situations. VLA-ATTC adaptively invokes test-time deliberation using an uncertainty-based cognitive clutch and a relative action critic for pairwise candidate selection. On LIBERO-LONG, the framework reduces PI0.5’s failure rate by over 50% while retaining practical control frequency.
Problem
Fast, intuitive VLA inference lacks deliberation and can produce suboptimal or catastrophic actions in complex or ambiguous situations.
Method
VLA-ATTC triggers test-time deliberation only when needed and uses a lightweight RAC model to select actions through iterative pairwise comparisons.
Results
VLA-ATTC reduces LIBERO-LONG’s SOTA failure rate by over 50%, boosts real-world success by 17.3%, and maintains a 20.8 Hz control frequency.
Takeaways & Limitations
Adaptive computation enables VLA models to allocate deliberation to difficult situations while preserving practical control frequency.
Abstract
from arXiv · showhide
Vision-Language-Action (VLA) models have demonstrated remarkable capabilities and generalization in embodied manipulation. However, their decision-making relies on a fast, instinctive process that lacks deliberation. This strategy often leads to suboptimal or catastrophic actions when facing complex or ambiguous scenarios that require greater consideration. In this paper, we introduce \textbf{VLA-ATTC}, a framework that endows VLA models with adaptive test-time compute (TTC). VLA-ATTC employs an uncertainty-based ``cognitive clutch'' to dynamically transition from reflexive execution to a TTC deliberation phase when necessary. During TTC phase, a novel \textbf{Relative Action Critic} (RAC) model identifies the optimal action from generated candidates via pairwise comparisons. This relative mechanism replaces unstable absolute value estimation, significantly simplifying the learning objective. Furthermore, we introduce an efficient sampling strategy to amortize computational costs and an automated data pipeline that curates preference pairs without manual annotation. On the LIBERO-LONG benchmark, VLA-ATTC reduces the failure rate of the SOTA model PI0.5 by over 50\%. We will open-source all the code and weights.
1 Introduction
VLA models generalize well but rely on fast inference that can fail in complex or ambiguous situations. VLA-ATTC adds adaptive test-time deliberation, a lightweight relative critic, and automated preference-data curation without fine-tuning the base model.
- Motivation: VLA models leverage pretrained world knowledge for diverse manipulation tasks, but fast intuitive inference can produce suboptimal or catastrophic failures in complex situations.The resulting capability gap motivates deliberative inference for difficult scenarios.
- Approach: VLA-ATTC addresses adaptive deliberation, computational overhead, and action-critic design through a cognitive clutch, efficient sampling, and a lightweight RAC model.The framework triggers deliberation only when uncertainty indicates it is necessary.
- Approach: The framework identifies optimal actions through iterative pairwise comparisons, avoiding unstable absolute action scoring.This design targets the accuracy and efficiency bottlenecks of prior parallel deliberation methods.
- Results: VLA-ATTC reduces LIBERO-LONG’s SOTA failure rate by over 50%, boosts real-world success by 17.3%, and maintains a 20.8 Hz control frequency.These results are reported alongside the framework’s automated preference-pair data pipeline.
2 Related Work
Prior VLA deliberation methods use sequential reasoning or parallel candidate selection, but they often require costly adaptation, computation, or unstable scoring. These limitations motivate more practical deliberation strategies.
- Sequential Deliberation: Sequential deliberation methods generate structured or free-form reasoning before actions, but require costly fine-tuning for auxiliary reasoning tasks.The cited approaches include ECoT, CoT-VLA, RoboMamba, PI0.5, ChatVLA, ChatVLA2, Hume, and OneTwoVLA.
- Parallel Deliberation: Parallel deliberation commonly uses Best-of-N sampling or Self-Consistency to select among multiple candidates.These paradigms are established across deliberative model settings and have been explored for VLAs.
- Parallel Deliberation: VLA parallel deliberation approaches face high computational cost, limited real-world world-model fidelity, or unstable absolute action scoring.Some methods use Monte Carlo Tree Search with a world model, while others use many candidates and a large external critic.
3 Preliminaries: VLA Inference Paradigm
VLA inference first encodes multimodal observations and instructions into a context representation, then decodes an action chunk stochastically. Because encoding dominates latency, multiple action candidates can be sampled relatively efficiently.
- Vision-Language Encoding: A pretrained VLM processes an image and language instruction to produce a multimodal context embedding containing scene and task information.This vision-language encoding uses a large Transformer forward pass and is computationally intensive.
- Action Decoding: An action head conditions on the encoded context to generate an action chunk.The action-generation stage is specialized to decode actions from the multimodal representation.
- Action Decoding: Diffusion or flow-matching models use an initial noise vector or random seed during stochastic action generation.The random seed is denoted by z in the described formulation.
- Efficiency: In PI0, action decoding takes 27ms of 86ms total inference time on RTX4090, so VLA-ATTC amortizes expensive pre-filling across multiple candidates.The VLM encoding stage constitutes most of the latency, while action decoding is significantly faster.
4 VLA-ATTC
VLA-ATTC adaptively switches between reflexive execution and test-time deliberation using uncertainty, then selects among parallel action candidates with a Relative Action Critic. The framework combines pairwise tournament selection, shared VLM computation, context-aware RAC inputs, and automated preference-pair curation.
- Uncertainty Quantification as Clutch: The Cognitive Clutch measures action-generation uncertainty and activates deliberation only when uncertainty exceeds a threshold.Low uncertainty leads to reflexive execution; high uncertainty triggers the slower deliberation path.
- TTC Deliberation Phase: When deliberation is triggered, the action head generates N candidate chunks in parallel using shared prefilled VLM context.Shared pre-filling makes the additional cost of generating candidates marginal.
- TTC Deliberation Phase: The framework replaces ambiguous absolute action scoring with pairwise preference comparisons organized as a single-elimination tournament.Each round pairs candidates, advances the preferred action, and repeats until one action remains.
- Relative Action Critic: The RAC receives two actions, their difference, and proprioceptive state, then predicts which action is preferred.Each input is projected by a dedicated MLP before entering the RAC Transformer.
- Relative Action Critic: The RAC fuses self-attention with raw and query cross-attention to use both VLM features and distilled task context.Learnable query tokens distill high-level semantic information during VLM prefilling, while a gated query branch controls contextual influence.
- Training and Data Curation: The RAC outputs a preference probability through an MLP classification head, trained with focal loss, while an automated pipeline creates preference pairs from existing action-generation data.The pipeline manipulates a flow-matching action head and requires no human intervention.
5 Experiments
Experiments evaluate VLA-ATTC across simulation and real-robot settings, testing effectiveness, component contributions, uncertainty estimation, preference-data validity, and efficiency. The framework improves success, identifies sparse difficult states, benefits from more candidates, and maintains high control frequency.
- Effectiveness: VLA-ATTC improves PI0 success on “Both pots on stove” from 40% to 58% and raises PI0’s average real-robot success rate by 17.3%.It also consistently surpasses RoboMonkey in the reported comparisons.
- Uncertainty threshold: Lowering the uncertainty threshold triggers more deliberation without significant performance gains, indicating that difficult states are sparse and the Cognitive Clutch identifies them effectively.The threshold τ controls when test-time deliberation is activated.
- Candidate scaling: Performance improves substantially with N = 4 candidates and continues rising as candidate count increases, while N = 16 is selected as an efficient choice under higher computational cost.The experiment varies the number of parallel candidate actions during deliberation.
- RAC ablations: Removing learnable weights, action-difference inputs, or learnable queries degrades LIBERO-LONG performance, with learnable-query removal causing the largest drop.The full RAC reaches 94% for PI0.5+ATTC and 90.8% for PI0+ATTC; removing learnable queries lowers these to 92.4% and 88%.
- Efficiency: Average control frequency decreases only from 23.3 Hz to 20.8 Hz with VLA-ATTC, compared with 1.5 Hz for RoboMonkey’s indiscriminate deliberation.The Cognitive Clutch limits expensive deliberation to a few timesteps.
6 Conclusion
VLA-ATTC replaces one-size-fits-all inference with uncertainty-triggered test-time deliberation and a lightweight Relative Action Critic for pairwise action selection. The framework improves robustness in complex scenarios while preserving the control frequencies required for robotics.
- VLA-ATTC dynamically triggers test-time deliberation through an uncertainty-based cognitive clutch.
- The Relative Action Critic selects actions through efficient pairwise comparisons.
- Adaptive computation improves decision-making robustness without sacrificing high robotic control frequencies.
- The framework supports agents that strategically allocate computational resources to the problem at hand.
A Verification of Semantic Preference Learning in RAC
The verification tests whether RAC learns semantic task compatibility rather than superficial generation quality. Using equally smooth candidates, RAC strongly preferred actions aligned with the visual task and instruction.
- The verification controls for low-level generation artifacts by using high-quality candidates produced with Nsteps = 10.
- The task-aligned candidate comes from the Stack Cubes expert policy, while the task-misaligned candidate comes from a Pour Water policy.
- 97.3% accuracy shows RAC preferred Stack Cubes actions over mismatched candidates.
- RAC evaluates compatibility among the proposed action, visual observation, and task instruction.