Research map / Vision-language and multimodal / Robot action and navigation
Robust Fine-tuning of Vision-Language-Action Robot Policies via Parameter Merging
Yajat Yadav, Zhiyuan Zhou, Andrew Wagenmaker, Karl Pertsch, Sergey Levine
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 |
|---|---|---|---|
| Interpolate finetuned and pretrained weights (RETAIN) | finetuned pi0-FAST-DROID · DROID whiteboard and plates, OOD | success rate: baselines 30-50% -> close to 80% (whiteboard), >60% (plates) | Improved |
| Merge weights with pretrained policy | finetuned policy · DROID and LIBERO, ID target task | success rate | No clear effect |
| Merge weights with pretrained policy | finetuned policy · Generalist tasks (44 DROID, 20 LIBERO) | success rate | Improved |
| Add co-finetuning before merging (RETAIN-co-FT vs RETAIN-task-FT) (ablation) | RETAIN · DROID and LIBERO, generalist and OOD | success rate | Improved |
| Use more pretraining data for the base policy (ablation) | RETAIN-co-FT · DROID plates, OOD | success gain over co-FT | Improved |
| Use more pretraining data for the base policy (ablation) | RETAIN-co-FT · DROID plates, ID | success rate | No clear effect |
| Merge only language-model parameters instead of all parameters (ablation) | RETAIN · LIBERO, OOD | success rate | No clear effect |
| Set alpha_l to 1 (no merging of language model) instead of 0.8 (ablation) | RETAIN modality-specific · LIBERO mugs-on-plates, OOD | success rate | Hurt |
| Set merging coefficient alpha too close to pretrained model (ablation) | RETAIN · LIBERO, OOD | success rate: near 0 performance | Hurt |
| Lower the learning rate or tune gradient steps in task-FT (ablation) | Task-FT · LIBERO, OOD | success rate | Hurt |
| Apply RETAIN in sequential skill acquisition instead of co-FT | DROID plates then whiteboard · ID and OOD | success rate | Improved |
| Use RETAIN with a less general base model (LIBERO) | RETAIN · LIBERO OOD vs DROID OOD | OOD gain over Co-FT | Mixed |
Closest papers by meaning
- Active Fine-Tuning of Multi-Task Policies (ICML 2025 · similarity 0.91)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning (ICLR 2025 · similarity 0.89)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.89)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.89)
- Zero-Shot Robotic Manipulation with Pre-Trained Image-Editing Diffusion Models (ICLR 2024 · similarity 0.88)
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.
