VLASelect: Selective Large-small Model Co-learning for Self-evolving VLA Agents

Le résumé fourni par la source

VLASelect Artifact Evaluation This repository contains the artifacts for the paper "VLASelect: Selective Large-small Model Co-learning for Self-evolving VLA Agents" (conditionally accepted by EuroSys'27). Checklist, Open Access Platforms and Downloads Artifact Evaluation Checklist (Available, Functional, Reproduced) An open access small machine and its Evaluation Report Open access of an academic cloud machine (CloudLab) and its Evaluation Report Docker Image: 147MB version (requiring further dependency installation by running dep-non-docker.sh) 33GB version (requiring no further dependency installation) Outline (Evaluation process/workflow and Reusability) 1. Artifact Overview 1.1 Introduction 1.2 Hardware/software Requirements and Dependencies 1.2.1 Hardware Requirements 1.2.2 Software Requirements 1.2.3 Get Source Code 1.2.4 Install Dependencies (if Docker can be installed) 1.2.5 Install Dependencies (if Docker cannot be installed) 1.2.6 Install Dependencies for Plotting Scripts 1.2.7 About Dataset 1.3 Treatment Measure for Unusual Behaviors 2. Evaluation Reproduction 2.1 One-click Reproduction 2.2 Step-by-Step Reproduction 2.2.1 Experiment 1: (Figure 7 in Section 5.2.1) Accuracy Under Tasks/Environment Changes 2.2.2 Experiment 2: (Figure 8 in Section 5.2.2) Accuracy Under Available Resource Changes 2.2.3 Experiment 3: (Figure 9 and Tables 2/3 in Section 5.3.1) Overheads Under The Same Accuracy 2.2.4 Experiment 4: (Figure 10 in Section 5.3.2) Time Breakdown of VLASelect's Modules 2.2.5 Experiment 5: (Figure 11 in Section 5.3.2) Training Time Breakdown in Each Workload 2.2.6 Experiment 6: (Figure 12 in Section 5.4) Design Choice Validation by Ablation 2.2.7 Experiment 7: (Discussion 1 in Section 5.5) Sim-to-real transfer 2.2.8 Experiment 8: (Discussion 2 in Section 5.5) ICL (In-Context Learning) 2.2.9 Experiment 9: (Discussion 3 in Section 5.5) Maximum Supported Model Size 2.2.10 Experiment 10: (Discussion 4 in Section 5.5) Applicability to Multi-Agent Scenarios 2.2.11 Experiment 11: (Discussion 5 in Section 5.5) Comparison with Alternative Model Scaling Techniques 2.2.12 Experiment 12: (Discussion 6 in Section 5.5) Comparison between Different Knowledge Exchange Granularities 2.2.13 Experiment 13: (Discussion 7 in Section 5.5) Forgetting on Previously Learned Environments/Tasks 2.2.14 Experiment 14: (Discussion 8 in Section 5.5) Applicability to MLP/CNN models 3. Reusability: Integrating VLASelect with VLA Models, Scaling Strategies, and Knowledge Exchange Granularities 3.1 Example 1: VLA-Adapter 3.1.1 Model Integration Interface 3.1.2 Integrating Different Scaling Strategies 3.1.3 Integrating Different Knowledge Exchange Granularities 3.2 Example 2: TinyVLA 3.2.1 Model Integration Interface 3.2.2 Integrating Different Scaling Strategies 3.2.3 Integrating Different Knowledge Exchange Granularities 3.3 Example 3: EdgeVLA 3.3.1 Model Integration Interface 3.3.2 Integrating Different Scaling Strategies 3.3.3 Integrating Different Knowledge Exchange Granularities

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.