Research map / Vision-language and multimodal / Robot action and navigation
VLA-Pro: Cross-Task Procedural Memory Transfer for Vision-Language-Action Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes VLA-Pro, which stores per-task LoRA adapters as procedural memories keyed by structured procedural states, then retrieves top-k and fuses them at inference to improve VLA cross-task generalization.
Problem
Vision-language-action models struggle to generalize to unseen tasks involving novel objects, scenes, and action patterns, even when these tasks are similar to training tasks.
Approach
Store each seen task as a (structured procedural state sequence, task-specific LoRA adapter) memory; at inference, extract the current procedural state with a VLM, retrieve the top-k most similar memories via action-aware field-weighted matching, and softmax-fuse their LoRA adapters into the VLA for the current action chunk.
Key results
- On RoboTwin unseen tasks, average success rises from 40.4% to 59.3% for pi0.5, 11.1% to 34.1% for RDT, and 17.0% to 30.0% for X-VLA.
- On 9 zero-shot RLBench tasks, VLA-Pro (k=2) reaches 20.9% average vs 13.8% for the pi0.5 baseline and 14.7% for AtomicVLA.
- Real-world held-out tasks with pi0.5 improve from 5.8% to 65.0% average success.
Surprising findings
- On RoboTwin, place_bell_behind rises from 0% to 100% for pi0.5, suggesting retrieval helps avoid dominant placement behaviors.
Limitations
- Performance depends on accuracy of VLM procedural state extraction and retrieval under ambiguous observations.
- Memory bank built from a limited set of seen tasks, restricting diversity and scalability.
- Retrieval and fusion of task-specific LoRA is similarity-based and cannot compose or abstract memories for highly novel tasks.
Experimental setup
Datasets: RoboTwin RLBench X-ICM RLBench split
Benchmarks: RoboTwin RLBench real-world UR7e tasks
Metrics: success rate mean reciprocal rank
Baselines: pi0.5 RDT X-VLA AtomicVLA Base LoRA
Builds on: X-ICM AtomicVLA
Closest papers by meaning
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- VLA-ATTC: Adaptive Test-Time Compute for VLA Models with Relative Action Critic Model (ICML 2026 · similarity 0.90)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.90)
- VLA-OS: Structuring and Dissecting Planning Representations and Paradigms in Vision-Language-Action Models (NeurIPS 2025 · similarity 0.90)
- Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization (NeurIPS 2025 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient 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.