Research map / Vision-language and multimodal / Robot action and navigation
Learning to Move Before Learning to Do: Task-Agnostic pretraining for VLAs
Junhao Shi, Siyin Wang, Xiaopeng Yu, Li Ji, Jingjing Gong, Xipeng Qiu
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 task-agnostic inverse-dynamics pretraining (TAP-20k) before behavior cloning | Qwen2.5-VL-3B VLA · SIMPLER WidowX | Avg-All success: 23.15% -> 33.32% | Improved |
| add task-agnostic inverse-dynamics pretraining | Qwen2.5-VL-3B VLA · SIMPLER WidowX | Avg-Partial success: 31.79% -> 45.82% | Improved |
| increase Stage 1 pretraining episodes from 8k to 14k to 20k (ablation) | TAP · SIMPLER WidowX | Avg-All success: 24.47% -> 30.21% -> 33.32% | Improved |
| scale Stage 1 pretraining steps (20k to 100k) with more Stage 2 finetuning (ablation) | TAP · SIMPLER WidowX | Avg-All success: stagnates near 18% at 20k steps; exceeds 30% at 100k steps | Improved |
| extend Stage 2 finetuning without enough Stage 1 pretraining (ablation) | TAP · SIMPLER WidowX | Avg-All success: stagnates near 18% | No clear effect |
| add task-agnostic pretraining on 30 hours of self-exploration (real world, push task, averaged) | WidowX policy with 200 expert demos · Real WidowX push pumpkin | success rate: 21% -> 61% | Improved |
| add task-agnostic pretraining (real world, carrot task, averaged) | WidowX policy with 200 expert demos · Real WidowX put carrot on plate | success rate: 9% -> 28% | Improved |
| task-agnostic pretraining under camera viewpoint perturbation | WidowX policy · Real WidowX Viewpoint Variation | success rate: 0% -> 15% (carrot), 0% -> 25% (pushing) | Improved |
| task-agnostic pretraining in clean standard setup versus large-scale OXE-pretrained NORA | TAP vs NORA · Real WidowX Standard Setup | success rate: TAP 75% vs NORA 85% (push); TAP 40% vs NORA 65% (carrot) | Hurt |
| add task-agnostic pretraining effect on convergence speed during Stage 2 (ablation) | TAP · SIMPLER WidowX | convergence speed: baseline saturates around 23%; pretrained exceed 30% | No clear effect |
Closest papers by meaning
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.91)
- VLMimic: Vision Language Models are Visual Imitation Learner for Fine-grained Actions (NeurIPS 2024 · similarity 0.90)
- FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.90)
- NS-VLA: Towards Neuro-Symbolic 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.
