🤖 machine learning

{\Omega}-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling

The paper introduces $\Omega$ -QVLA, a training-free post-quantization framework that enables robust W4A4 uniform quantization for entire Vision-Language-Action models by combining composite SVD-Hadamard rotation with per-step DiT activation scaling, achieving state-of-the-art task success rates and significant memory reduction on both benchmarks and real-world manipulation tasks.

Original authors: Xinyu Wang, Mingze Li, Sicheng Lyu, Dongxiu Liu, Kaicheng Yang, Ziyu Zhao, Yufei Cui, Xiao-Wen Chang, Peng Lu

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xinyu Wang, Mingze Li, Sicheng Lyu, Dongxiu Liu, Kaicheng Yang, Ziyu Zhao, Yufei Cui, Xiao-Wen Chang, Peng Lu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant, multi-billion-dollar robot brain. This brain is a "Vision-Language-Action" (VLA) model. It can see the world through cameras, understand complex spoken instructions, and figure out exactly how to move its robotic arms to do a task.

The problem? This brain is huge. It's like trying to run a massive, high-end movie theater on a tiny, battery-powered handheld game console. It takes up too much memory and is too slow to run on actual robots.

To fix this, scientists usually try to "shrink" the brain by compressing its numbers (a process called quantization). Think of it like converting a high-definition 4K movie into a low-resolution MP4 to save space.

The Big Problem:
Previous attempts to shrink these robot brains had a major flaw. They would compress the "thinking" part (the language model) but leave the "moving" part (the action head) at full, high-quality resolution. Why? Because everyone believed the "moving" part was too delicate. They thought that if you compressed it too much, the robot would start shaking, jerking, or dropping things because the numbers got too "fuzzy."

The Solution: Ω-QVLA
This paper introduces a new method called Ω-QVLA. It's the first tool that successfully shrinks both the thinking and moving parts of the robot brain to the same tiny size (called W4A4) without the robot falling apart.

Here is how they did it, using two creative tricks:

1. The "Spinning Shuffle" (Composite Rotation)

Imagine you have a deck of cards where a few cards are incredibly heavy (outliers) and the rest are light. If you try to pack them into a small box, the heavy cards crush the others, and the box breaks.

In the robot brain, some data channels are "heavy" (they have huge numbers) while others are "light."

The Old Way: Just squish them all together. The heavy cards crush the box.
The Ω-QVLA Way: They use a special "shuffle" (a mix of math called SVD and Hadamard rotation). Imagine spinning the deck of cards and shuffling them so that the weight of the heavy cards is spread out evenly across the whole deck. Now, no single card is heavy enough to break the box. This allows them to compress the data much more aggressively without losing the "shape" of the information.

2. The "Step-by-Step Volume Control" (Per-Step Scaling)

The robot's "moving" part works like a video editor smoothing out a video frame by frame (a process called denoising).

The Problem: The "volume" (or intensity) of the data changes wildly from the first frame to the last frame. If you set one fixed volume level for the whole video, the beginning might be too quiet, and the end might be too loud and distorted.
The Ω-QVLA Way: They created a "dynamic volume control" that adjusts the sensitivity for every single step of the video. Instead of using one static setting, the system listens to the data at each specific moment and adjusts the compression scale accordingly. This prevents the robot from getting confused when the data gets too loud or too quiet during its movement.

The Results

The team tested this on two famous robot brains (Pi 0.5 and GR00T N1.5) using a simulation called LIBERO and real-world robot arms.

In Simulation: The compressed robots performed just as well as the giant, uncompressed ones. In fact, on one test, the tiny compressed robot actually did a better job than the original!
In the Real World: When they put the compressed robot on a real table to pick up cups, fold towels, and move blocks, it moved smoothly and accurately.
The Competition: Other methods that tried to do this (QuantVLA) made the robots jerk around and fail the tasks. Ω-QVLA kept them smooth.
Space Saved: They reduced the memory needed by 71%. That's like turning a 4GB movie file into a 1GB file without losing the plot.

The Bottom Line

This paper proves that the "moving" part of a robot brain isn't too fragile to be compressed. By using a smart "shuffle" to spread out the heavy data and a "step-by-step" adjustment to handle changing volumes, they can shrink these massive robot brains down to a size that fits on a real robot, making them faster, cheaper, and ready for the real world.

Technical Summary: Ω-QVLA

Problem Statement

Vision-Language-Action (VLA) models integrate perception, reasoning, and control into a single policy, typically utilizing a large language model (LLM) backbone and a diffusion-based action head (DiT). While promising for generalist robot control, these models are prohibitively expensive to deploy on-device due to their multi-billion parameter counts and memory requirements.

Existing post-training quantization (PTQ) efforts have failed to provide a complete solution for VLA models. Prior work typically compresses the LLM backbone while leaving the DiT action head at full precision (FP16) or employs mixed-precision schemes. This hesitation stems from the belief that the DiT action head is inherently too sensitive to uniform low-bit quantization. Unlike language generation, where quantization errors might be imperceptible, errors in the action head produce continuous control signals that directly interface with physical actuators. In a closed-loop setting, quantization-induced perturbations are amplified by contact forces and physical dynamics, leading to cumulative error propagation and task failure. Consequently, no prior method has successfully quantized the full DiT action head of a VLA model to a uniform low-bit precision.

Methodology: Ω-QVLA

The authors propose Ω-QVLA, the first training-free PTQ framework capable of compressing both the language backbone and the entire DiT action head of a VLA model to a uniform W4A4 (4-bit weights, 4-bit activations) precision. The framework eliminates the need for mixed-precision allocation through two core innovations:

1. Composite SVD·Hadamard Rotation

To address channel-level energy imbalance and activation outliers, the authors introduce a two-level rotation strategy applied before quantization:

SVD Rotation: A singular value decomposition (SVD) is computed on the weight matrix $W = U\Sigma V^\top$ . The rotation matrix $R_{SVD} = U$ is applied to equalize the row-wise energy of the weights. This transforms the weight space such that the energy of each row is directly controlled by the singular spectrum, removing coordinate-induced energy concentration.
Hadamard Rotation: Since SVD rotation is derived solely from weights, it does not guarantee uniformity in the activation space. To address this, the authors compose the SVD rotation with a normalized Hadamard matrix $H$ . The composite rotation $R = R_{SVD} \cdot H$ acts as a uniform mixing matrix that redistributes the energy of any dominant channel uniformly across all output channels.
Block-wise Implementation: To avoid the computational and memory overhead of a dense $C_{in} \times C_{in}$ rotation matrix, the method approximates the rotation in blocks. Channels are reordered via a zigzag permutation based on weight column norms to ensure that high-norm and low-norm channels are distributed evenly across blocks, preventing outlier concentration within a single partition.

2. Per-step DiT Activation Scaling

Standard static scaling is suboptimal for DiT models because activation magnitudes fluctuate significantly across the multi-step denoising process.

Dynamic Range Drift: The authors observe that activation statistics drift across denoising steps, particularly in layers receiving time-conditioned inputs (e.g., QKV projections following AdaLayerNorm).
Per-step Calibration: Instead of a single static scale, Ω-QVLA employs a precomputed, per-step, per-layer, and per-channel scale table. This table is derived from calibration trajectories and retrieved during inference to absorb dynamic-range drift. This allows the quantizer to adapt to the shifting activation ranges at each specific timestep.

Key Contributions

Unified W4A4 Quantization: Ω-QVLA is the first training-free framework to uniformly quantize the entire VLA stack (LLM backbone + DiT action head) to W4A4, challenging the assumption that the action head cannot survive uniform low-bit quantization.
Composite Rotation Mechanism: The introduction of a composite SVD·Hadamard rotation that simultaneously equalizes per-channel weight energy and diffuses residual activation outliers, addressing the limitations of using SVD or Hadamard transformations in isolation.
Per-step Calibration: The development of a per-step activation scaling mechanism specifically for DiT models to handle timestep-dependent dynamic-range drift, which is critical for maintaining stability in iterative denoising processes.

Experimental Results

The framework was evaluated on the LIBERO benchmark using two state-of-the-art VLA policies: Pi 0.5 and GR00T N1.5.

Simulation Performance:
- Pi 0.5: Achieved a 98.0% average task success rate under W4A4 quantization, slightly exceeding its FP16 reference (97.1%).
- GR00T N1.5: Achieved an 87.8% average success rate, matching its FP16 reference (87.0%).
- Comparison: These results significantly outperform prior W4A4 full-quantization baselines (e.g., SmoothQuant, DuQuant, QuantVLA), which often failed to maintain performance when quantizing the DiT attention layers. For instance, on Pi 0.5, QuantVLA achieved only 82.0% success, while Ω-QVLA reached 98.0%.
- Memory Efficiency: The framework reduced the static memory footprint by 71.3% compared to FP16, with negligible metadata overhead (~90 MB for Pi 0.5).
Real-World Manipulation:
- Experiments were conducted on a dual-arm ARX R5 platform with five bimanual manipulation tasks (e.g., Pick Cup, Fold Towel).
- Performance: Ω-QVLA achieved an average progress score of 51.0, surpassing the FP16 baseline (49.6). In contrast, the baseline QuantVLA collapsed to 25.0.
- Qualitative Analysis: Real-world tests confirmed that Ω-QVLA produces smooth, accurate manipulation trajectories. Baseline methods exhibited jerky end-effector movements and frequent spikes in action dimensions, leading to task failures over long horizons.

Significance and Claims

The paper claims that Ω-QVLA overturns the prevailing belief that the DiT action head is too sensitive to be uniformly quantized. By directly addressing the sources of instability—channel-level energy imbalance and denoising-step dynamic-range drift—the authors demonstrate that a unified W4A4 precision is sufficient for high-performance VLA deployment.

The authors emphasize that their approach requires no retraining and eliminates the complexity of mixed-precision allocation. The results suggest that effective VLA compression requires holistic, action-aware optimization rather than the direct transfer of language-centric quantization techniques. While the current evaluation is limited to DiT-based backbones and specific benchmarks, the work establishes a new baseline for on-device robot control, enabling the deployment of large-scale VLA policies on hardware with strict memory and latency constraints.