(논문 요약) Mixture-of-Agents Enhances Large Language Model Capabilities (Paper)

핵심 내용

• vision-language-action 모델
• 970k real-world robot demonstrations (Open X-Embodiment dataset)
• Parameter Efficient FineTuning 으로 new task 학습

• 조금 움직이고 다시 조금 움직이며 동작 (데모 동영상은 배속)
  

• archiectue
  • 600M visual encoder (SigLIP and DinoV2 - channelwise-concat)
  • 2-layer MLP projector
  • 7B Llama2
    
• robot action
  • discretized action space (1st and 99th quantile for each variable in training data)
• data curation
  • Open X dataset: >70 individual robots, >2M robot trajectories
  • Open X-Embodiment dataset: 다음 기준으로 Open X 에서 샘플
    • coherent input and output space across all training
    • balanced mix of embodiments, tasks, and scenes
• 결과