Reproduce “EveryDayVLA: A Vision-Language-Action Model for Affordable Robotic Manipulation” on OpenResearch

About this paper

Samarth Chopra, Alex McMoil, Ben Carnovale, Evan Sokolson, Rajkumar Kubendran, Samuel Dickerson

Researchers at the University of Pittsburgh developed EveryDayVLA, a full-stack system combining a $300 6-DOF robotic manipulator with an adaptive Vision-Language-Action (VLA) model for affordable and robust manipulation. The system achieved a 91.4% success rate on the LIBERO simulation benchmark and demonstrated a 49% higher real-world success rate for in-distribution tasks, along with 34.9% better out-of-distribution generalization compared to existing methods.

Key insights

• Accessible, low-cost robotic hardware can achieve high performance when paired with robust and intelligently designed Vision-Language-Action models.
• Dynamically adapting the planning horizon based on real-time model uncertainty, specifically the disagreement between discrete and continuous action predictions, significantly improves VLA robustness and generalization.
• Jointly training VLA models on both continuous action trajectories and discrete action tokens provides a richer learning signal and enables more resilient performance in diverse scenarios.

Summary by alphaXiv.