Recoleta
GitHub
Source note

Ω-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling

Authors
Xinyu Wang, Mingze Li, Sicheng Lyu, Dongxiu Liu, Kaicheng Yang, Ziyu Zhao, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionVLA QuantizationRobot Policy CompressionDiffusion Action HeadPost Training Quantization
Open arXivMarkdown source

Summary

Ω-QVLA is a training-free post-training quantization method for VLA robot policies. It quantizes both the language backbone and DiT action head to uniform W4A4 while keeping LIBERO success rates near FP16.

Problem

  • VLA models such as Pi 0.5 and GR00T N1.5 use billion-parameter backbones plus diffusion action heads, which makes robot-side deployment expensive in memory and compute.
  • Existing PTQ methods often leave DiT attention or the action head at higher precision because action errors affect continuous robot control.
  • Uniform W4A4 quantization matters because equal 4-bit weights and activations can reduce memory and support lower-precision compute without mixed-precision handling.

Approach

  • Applies orthogonal rotations before quantization so each linear layer can be computed as (X R)(R^T W), then quantizes the rotated tensors.
  • Uses SVD rotation to reduce weight channel energy imbalance; the paper reports weight row-norm spread dropping from 26× to 6× in an example layer.
  • Adds a Hadamard rotation after SVD to spread activation outliers; the same example reports activation spread dropping from 20× to 1.6×.
  • Uses block-wise SVD-Hadamard rotation with block size 64 plus zigzag channel permutation by weight norm to control cost.
  • Calibrates DiT activation scales per layer, denoising step, and channel over 10 unlabeled trajectories and 8 Euler denoising steps.

Results

  • On LIBERO, Ω-QVLA W4A4 reaches 98.0% average success on Pi 0.5 versus 97.1% FP16; task scores are Goal 100.0, Spatial 99.0, Object 97.0, Long 96.0.
  • On LIBERO, Ω-QVLA W4A4 reaches 87.8% average success on GR00T N1.5 versus 87.0% FP16; task scores are Goal 91.0, Spatial 86.0, Object 92.0, Long 82.0.
  • Compared with W4A4 full QuantVLA on Pi 0.5, average success rises from 82.0% to 98.0%, and Long rises from 56.0% to 96.0%.
  • Compared with W4A4 full baselines on GR00T N1.5, Ω-QVLA scores 87.8% average versus SmoothQuant 84.0%, DuQuant 70.0%, and QuantVLA 69.8%.
  • The method reports a 71.3% static memory-footprint reduction.
  • In real-world Pi 0.5 W4A4 tests on an ARX R5 dual-arm robot, Ω-QVLA scores 51.0 average progress across 5 tasks versus 49.6 for FP16 Pi-0.5 Base and 25.0 for QuantVLA.

Link