Research map / Vision-language and multimodal / Robot action and navigation
VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers
Yating Wang, Haoyi Zhu, Mingyu Liu, Jiange Yang, Hao-Shu Fang, Tong He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace OpenVLA action tokens with Residual VQ-VAE tokenizer (trained on ManiSkill+RLBench) | OpenVLA · LIBERO-90 | success rate (%): 73.53 -> 80.98 | Improved |
| train tokenizer on ManiSkill only instead of ManiSkill+RLBench (ablation) | VQ-VLA tokenizer · LIBERO-90 | success rate (%): 80.98 -> 14.38 | Hurt |
| use temporal conv encoder/decoder instead of MLP (ablation) | Residual VQ-VAE · LIBERO-10 | success rate (%): 53.4 -> 60.0 | Improved |
| use temporal conv instead of MLP (ablation) | Residual VQ-VAE · LIBERO-GOAL | success rate (%): 72.6 -> 75.2 | Improved |
| train tokenizer on all LIBERO instead of LIBERO-10 only (ablation) | Conv Residual VQ-VAE · LIBERO-10 | success rate (%): 54.0 -> 60.0 | Improved |
| train tokenizer on all LIBERO instead of LIBERO-GOAL only (ablation) | Conv Residual VQ-VAE · LIBERO-GOAL | success rate (%): 72.4 -> 75.2 | Improved |
| add ManiSkill synthetic data to tokenizer training (ablation) | VQ_O+L · Real-world: Flip the pot upright | success rate (%): 45.0 -> 60.0 | Improved |
| add ManiSkill synthetic data to tokenizer training (ablation) | VQ_O+L · Real-world: Put the toy into the drawer | success rate (%): 10.0 -> 25.0 | Improved |
| add LIBERO data to tokenizer training (VQ_O to VQ_O+L) (ablation) | VQ-VLA · Real-world short-horizon tasks | average success rate (%): increased by only 0.5% | No clear effect |
| train tokenizer on LIBERO only (VQ_L) vs real+LIBERO (ablation) | VQ-VLA · Real-world tasks | success rate (%): 55.0 vs 45.0 (Flip pot) | No clear effect |
| use VQ-based action chunking instead of autoregressive chunking (ablation) | OpenVLA · LIBERO-90 | success rate (%): 66.53 -> 86.61 | Improved |
| use autoregressive 5-action chunking in OpenVLA (ablation) | OpenVLA · Real-world: Put the toy into the basket | success rate (%): 20.0 -> 0.0 | Hurt |
Later papers that compared with this method
2 later accepted papers report VQ-VLA as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- M2Tok: Multi-head Multi-codebook Discrete Action Tokenization for Vision-Language-Action Models (ECCV 2026 · better on 9 of 9 shared results)
- MoEActok: A MoE-based Action Tokenizer for Vision-Language-Action Models (CVPR 2026 · better on 5 of 7 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ActQuant: Sub-4-bit Action-Guided Quantization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- RotVLA: Rotational Latent Action for Vision-Language-Action Model (NeurIPS 2026 · similarity 0.89)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- ATI-VLA: Action-Centric Predictive Vision–Language–Action Models via Actionable Alignment Then Adaptive Injection (NeurIPS 2026 · similarity 0.89)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
