Source-linked AI summary
QVLA: Not All Channels Are Equal in Vision-Language-Action Model's Quantization
Yuhao Xu, Yantai Yang, Zhenyang Fan, Yufan Liu, Yuming Li, Bing Li, Zhipeng Zhang
TL;DR
VLA quantization lacks methods that account for action-space sensitivity, making uniform-bit strategies ill-suited to action-driven models. QVLA addresses this with per-channel sensitivity-guided allocation that unifies quantization and pruning, and evaluations show stronger performance than adapted LLM/MLLM methods.
Problem
VLA quantization requires an approach beyond uniform bit-width because different parameters and channels have heterogeneous sensitivity to quantization noise.
Method
QVLA uses action-space sensitivity to rank channels and a greedy global allocation procedure that jointly supports mixed-precision quantization and 0-bit pruning.
Results
QVLA achieves lower action errors and higher task success rates than adapted LLM/MLLM methods at equivalent average bit-widths, while channel-wise quantization reaches 76.8% versus 74.9% for layer-wise quantization at INT8.
Takeaways & Limitations
Action-space alignment and channel-wise allocation provide a foundation for compressing VLA models for embodied AI.
Takeaways & Limitations
The sensitivity analysis has a stated limitation involving single-step error metrics such as s(b).
Abstract
from arXiv · showhide
The advent of Vision-Language-Action (VLA) models represents a significant leap for embodied intelligence, yet their immense computational demands critically hinder deployment on resource-constrained robotic platforms. Intuitively, low-bit quantization is a prevalent and preferred technique for large-scale model compression. However, we find that a systematic analysis of VLA model's quantization is fundamentally lacking. We argue that naively applying uniform-bit quantization from Large Language Models (LLMs) to robotics is flawed, as these methods prioritize passive data fidelity while ignoring how minor action deviations compound into catastrophic task failures. To bridge this gap, we introduce QVLA, the first action-centric quantization framework specifically designed for embodied control. In a sharp departure from the rigid, uniform-bit quantization of LLM-based methods, QVLA introduces a highly granular, channel-wise bit allocation strategy. Its core mechanism is to directly measure the final action-space sensitivity when quantizing each individual channel to various bit-widths. This process yields a precise, per-channel importance metric that guides a global optimization, which elegantly unifies quantization and pruning (0-bit) into a single, cohesive framework. Extensive evaluations on different baselines demonstrate the superiority of our approach. In the LIBERO, the quantization version of OpenVLA-OFT with our method requires only 29.2% of the original model's VRAM while maintaining 98.9% of its original performance and achieving a 1.49x speedup. This translates to a 22.6% performance improvement over the LLM-derived method SmoothQuant. Our work establishes a new, principled foundation for compressing VLA models in robotics, paving the way for deploying powerful, large-scale models on real-world hardware. Code will be released.
1 INTRODUCTION
VLA models improve embodied intelligence but remain difficult to deploy because of high memory and latency demands. QVLA addresses the lack of VLA-specific quantization analysis with action-centric, channel-wise bit allocation that unifies quantization and pruning.
- Motivation: 7B VLA models often exceed 14 GB in half-precision, creating a deployment barrier for resource-constrained robotic platforms.Inference on an NVIDIA Jetson AGX Orin can take several hundred milliseconds per action.
- Quantization Gap: Existing quantization methods lack systematic analysis of their unique impacts and trade-offs when applied to VLA models.
- Quantization Gap: LLM- and MLLM-derived methods preserve text perplexity or visual fidelity, but VLA outputs are continuous action values that directly interface with the physical world.
- Quantization Gap: Outlier-centric and module-level mixed-precision approaches miss cross-modal, action-decoding, and intra-layer channel sensitivity in VLA models.Long-horizon tasks can amplify small initial errors, while channels within the same layer contribute differently to final outputs.
- QVLA: QVLA anchors quantization in action space, allocates bits per channel, and unifies weight quantization with pruning through 0-bit allocation.The framework is presented as the first systematic VLA-specific quantization analysis and action-centric compression method.
2 RELATED WORK
VLA research includes both model architectures for mapping vision and language to actions and quantization methods for reducing deployment cost. Existing quantization work centers on PTQ, QAT, outlier mitigation, rescaling, and protection of salient weights.
- Vision-Language-Action Models: VLA methodologies diverge according to action decoding: some discretize continuous actions into sequence-to-sequence prediction, while others follow a different decoding strategy.
- Quantization Methods: Quantization reduces memory footprint and computational latency by representing weights and activations with low-bit integers, while seeking to limit accuracy degradation.
- Quantization Methods: Post-Training Quantization uses a small calibration set without retraining, whereas Quantization-Aware Training simulates quantization during training.
- Quantization Methods: Recent methods address outliers and activation difficulties through rescaling, salient-weight preservation, and weight-activation smoothing.Examples include SmoothQuant, AWQ, and OmniQuant.
3 METHOD
QVLA formulates VLA quantization around action-space fidelity and allocates precision at channel granularity. Its sensitivity-guided optimization accounts for heterogeneous module and channel impacts, including long-horizon error accumulation, while combining quantization with pruning.
- Preliminaries: VLA quantization minimizes divergence between the original and quantized policy action distributions rather than only approximating internal representations.The formulation treats the policy’s action output as the central preservation target.
- Sensitivity Analysis: Projectors and action heads are more sensitive than vision encoders, because perturbations near action decoding propagate directly into the output action distribution.This module-level disparity motivates adaptive precision across the architecture.
- Sensitivity Analysis: Channel sensitivity is heterogeneous within layers, so global or per-layer uniform bit allocation cannot preserve all channels equally.A channel denotes an output channel in a convolutional layer or a row in a linear-layer weight matrix.
- Action-Space Sensitivity: QVLA measures each channel’s action-space sensitivity by quantizing it alone at 0, 2, 4, 8, or 16 bits while keeping other parameters full precision.The single-step metric uses the expected squared L2 action deviation; cumulative sensitivity captures episode-level error accumulation and correlates more strongly with task success.
- Optimal Bit Allocation: QVLA assigns channel bit-widths with greedy demotion, repeatedly lowering the least sensitive channels through 16→8→4→2→0 until the target average bit budget is met.The method uses a sensitivity-to-bit ratio for marginal error per bit saved and applies regularization to the final pruning stage; sorting costs O(C log C).
4 EXPERIMENTS
Experiments evaluate QVLA across weight-activation, weight-only, channel-wise, temporal, pruning, and real-world settings. Across these evaluations, QVLA preserves task performance while reducing memory and improving inference efficiency.
- Weight-activation quantization: QVLA retains 99.3% of OpenVLA’s full-precision performance under W4A4, using 28.2% of memory and achieving a 1.47× inference speedup.The accuracy drop is 0.5%, compared with 13.3% for SmoothQuant and 3.2% for OmniQuant.
- Weight-only quantization: Under W4A16 weight-only quantization, QVLA incurs zero performance loss on OpenVLA, while AWQ’s average success rate drops by 4.7%.For OpenVLA-OFT, QVLA’s degradation is 0.1% and 0.4%, compared with AWQ losses of 2.0% and 4.5%.
- Channel-wise quantization: Channel-wise quantization preserves OpenVLA’s 76.5% full-precision success rate at INT4 and reaches 76.8% at INT8, whereas layer-wise quantization falls to 74.9% at INT8.The comparison is reported on LIBERO.
- Temporal error accumulation: QVLA’s 8-bit method maintains lower cumulative action error than the uniform 8-bit baseline, with the gap widening over longer time horizons.Cumulative error grows faster for 4-bit quantization than for 8-bit methods.
- Pruning and uniform bits: With an overall INT8 budget, channel-wise gating plus pruning reduces memory to 7.0 GB and achieves a 76.8% average success rate, while uniform 8-bit quantization reaches 74.6%.The channel-wise candidate set without pruning already matches the full-precision baseline at 76.7% versus 76.5%.
5 CONCLUSION
The paper concludes that uniform-bit quantization transferred from LLMs and MLLMs is ill-suited to action-driven VLA models. QVLA instead uses action-space sensitivity for adaptive per-channel allocation and unifies quantization with pruning.
- The paper presents the first systematic analysis of quantization challenges specific to VLA models.
- Uniform-bit quantization can significantly degrade VLA performance because these models are uniquely sensitive to quantization noise.
- QVLA uses a per-channel action-space sensitivity metric and a global greedy algorithm to unify quantization and pruning.
- Evaluations on OpenVLA and OpenVLA-OFT show that QVLA outperforms conventional methods, reduces action errors, and can surpass full-precision task success rates.
C IMPACT OF THE GATES RATIO
Under an overall INT8 budget, performance depends strongly on the proportion of parameters assigned 8-bit precision. The best reported success rate occurs when most parameters are quantized to 8-bit.
- 76.3% success rate is achieved at the optimal gate ratio under an overall INT8 budget.Performance systematically degrades as the proportion of parameters quantized to 8-bit decreases.
D EXPANDED EVALUATION ON UNIVLA
The expanded evaluation tests AutoQVLA on UniVLA-7B and the CALVIN benchmark. The results support transfer of action-centric channel sensitivity across diverse VLA architectures and low-bit settings.
- The CALVIN evaluation reports task success rates by sequence length for compressed models under W8A16.
- Action-centric channel sensitivity generalizes to UniVLA-7B, whose internal structure uses task-centric latent action decoding.
- 95.1% average success rate is achieved by AutoQVLA under UniVLA-7B W4A16 quantization, compared with 92.6% for AWQ.AutoQVLA also produces the lowest memory footprint among the tested quantization methods.
- AutoQVLA consistently surpasses GPTQ, AWQ, SmoothQuant, and OmniQuant across various low-bit configurations on LIBERO.
E EVALUATION ON THE CALVIN BENCHMARK
On CALVIN, AutoQVLA (W8A16) preserved long-horizon task success and sequence stability with negligible performance degradation despite the benchmark’s more intricate planning and interaction challenges.
- AutoQVLA (W8A16) on OpenVLA-OFT exhibited negligible performance degradation on CALVIN.CALVIN evaluates more intricate sequence planning and interaction challenges than LIBERO.
F CALIBRATION SET DETAILS AND SIZE ABLATION
The LIBERO calibration set uses sampled training trajectories for action-space sensitivity analysis, with 512 trajectories achieving the best reported average success rate and neighboring sizes remaining stable.
- 512 LIBERO training trajectories produced the highest average success rate of 97.0% for OpenVLA-OFT (W8A16).The trajectories are sampled from combined task training data and used for sensitivity analysis and bit allocation.
- 96.8% and 96.7% average success rates at 256 and 1024 trajectories indicate only marginal degradation around the 512-trajectory setting.The ablation therefore reports stable performance across neighboring calibration-set sizes.
G THEORETICAL ANALYSIS OF ACTION-SPACE SENSITIVITY APPROXIMATION
The theoretical analysis defines action-space sensitivity as output-action MSE from channel quantization and approximates it using a first-order Jacobian-based formulation under a uniform quantization-noise model.
- Action-space sensitivity is defined as the action-space MSE caused by quantizing layer l, channel c to bit-width b.The action mapping A converts channel output perturbations into final actions, while quantization noise represents the channel perturbation.
- A first-order Taylor expansion approximates action deviation using the Jacobian of the action with respect to the channel output.The approximation assumes local linearity for small perturbations.
- Under zero-mean isotropic uniform quantization noise, the expected quadratic form simplifies through the trace identity.This assumption supports the analytical sensitivity approximation.
- The resulting first-order proxy incorporates the squared Frobenius norm of the action Jacobian.The derivation provides an efficient proxy for estimating action-space sensitivity.
H QUALITATIVE ANALYSIS OF QUANTIZED POLICY BEHAVIOR
Qualitative rollouts show that quantization differences emerge most strongly in fine-grained manipulation and recovery, while cumulative sensitivity rankings largely agree with single-step rankings.
- Quantization schemes mainly differ in action precision and stability when the full-precision model completes the task.Competing quantized-model failures concentrate around grasping, stable contact, and final placement.
- Quantized models generally show weaker robust re-localization than the full-precision baseline after initial misses or accumulated errors.The full-precision model typically produces more reasonable fallback motions during recovery.
- AutoQVLA’s quantitative advantage is attributed to preserving high-fidelity control signals where small deviations compound into physical errors.This connects the qualitative control behavior to the reported performance differences.
- 80% of sampled channels received closely matching ranks from cumulative and single-step sensitivity metrics.The scatter comparison used 1,000 randomly sampled channels, with diagonal proximity indicating consistent rankings.