Research map / Vision-language and multimodal / Robot action and navigation
MetaVLA: Unified Meta Co-Training for Efficient Embodied Adaptation
Chen Li, Zhantao Yang, Han Zhang, Fangyi Chen, Chenchen Zhu, Anudeepsekhar Bolimera, Marios Savvides
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 |
|---|---|---|---|
| add auxiliary GR00T tasks to naive multi-task SFT (ablation) | SFT-4LIBERO (OpenVLA) · LIBERO average | success rate (%): 76.2 -> 8.6 | Hurt |
| train naive SFT with auxiliary tasks for more steps (187.5K vs 75K) (ablation) | SFT-4LIBERO+5single+1bimanual · LIBERO average | success rate (%): 8.6 -> 14.5 | Mixed |
| add Meta-Action-Reasoner context module with LIBERO context only (ablation) | SFT-4LIBERO (OpenVLA) · LIBERO average | success rate (%): 76.2 -> 77.8 | Improved |
| add six auxiliary tasks to the MetaVLA context bank (ablation) | MetaVLA · LIBERO average | success rate (%): 77.8 -> 79.3 | Improved |
| replace exotic auxiliary tasks with tasks already in OpenVLA pretraining (ablation) | MetaVLA · LIBERO average | success rate (%): 77.8 -> 74.4 | Hurt |
| add stochastic latent with KL regularization (ablation) | MetaVLA · LIBERO-Long | success rate (%): 55.3 -> 53.0 | Hurt |
| increase context batch size from 4 to 32 (ablation) | MetaVLA · LIBERO average | success rate (%): 73.2 -> 77.8 | Improved |
| apply MetaVLA to NORA-Long without auxiliary tasks (ablation) | NORA-Long · LIBERO average | success rate (%): 87.3 -> 90.3 | Improved |
| add auxiliary tasks to naive SFT on NORA-Long (ablation) | NORA-Long-SFT-4LIBERO · LIBERO average | success rate (%): 87.3 -> 66.4 | Hurt |
| add auxiliary tasks to MetaVLA on NORA-Long (ablation) | MetaVLA-NORA-Long · LIBERO average | success rate (%): 90.3 -> 91.8 | Improved |
| train MetaVLA separately per suite (MetaVLA-EACH) instead of one co-trained model (ablation) | MetaVLA · LIBERO-Long, 240K steps | success rate (%): 51.8 -> 55.8 | Improved |
| unify four suite models into one MetaVLA model | MetaVLA · LIBERO training cost | training steps: 240K -> 75K | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TraceVLA — TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies (ICLR 2025) · better on 1 of 1 shared result
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
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
- UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent (ICML 2025 · similarity 0.89)
- ModularAgent: A Task-Aware Modular Framework for Joint Optimization of Multimodal Large Language Models and World Models (CVPR 2026 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents (ICML 2025 · similarity 0.89)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (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.
