Research map / Vision-language and multimodal / Robot action and navigation
VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model
Yanjiang Guo, Tony Lee, Lucy Xiaoyang Shi, Jianyu Chen, Percy Liang, Chelsea Finn
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 |
|---|---|---|---|
| finetune the world model on expert rollouts and then online policy rollouts (ablation) | Ctrl-World · action replay, 256 clips | FVD: 225.13 -> 99.98 -> 64.12 | Improved |
| add online rollout data (success and failure) to world model finetuning (ablation) | Ctrl-World + expert rollout · interaction event confusion matrix, 50 clips | false positives: 11 -> 1 | Improved |
| add online rollout data to world model finetuning (ablation) | Ctrl-World + expert rollout · interaction event confusion matrix, 50 clips | true positives: 28 -> 26 TP, 9 -> 19 TN | Mixed |
| finetune world model on expert demonstrations only (ablation) | Ctrl-World · action replay, qualitative | interaction outcome fidelity | Hurt |
| add world-model synthetic rollouts with reward filtering on top of real rollouts | pi0.5 policy · DROID five tasks, multi-task | success rate: +11.6% over baseline | Improved |
| use DSRL noise-space optimization instead of VLAW | pi0.5 policy · DROID five tasks, multi-task | success rate | No clear effect |
| filtered behavior cloning on real successful rollouts only | pi0.5 policy · DROID five tasks, multi-task | success rate | Improved |
| reduce synthetic rollouts from 500 to 250 (ablation) | VLAW · Drawing task, 1st round | success rate (%): 68 -> 60 | Hurt |
| add real-world success trajectories to policy finetuning with synthetic data (ablation) | VLAW · Drawing task, 1st round | success rate (%) | Improved |
| threshold the reward model's yes probability | VLM reward model · synthetic trajectory filtering | false-positive trajectories | Improved |
Closest papers by meaning
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Knowledge Insulating Vision-Language-Action Models: Train Fast, Run Fast, Generalize Better (NeurIPS 2025 · similarity 0.90)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.90)
- Beyond World-Frame Action Heads: Motion-Centric Action Frames for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
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.
