Researchers have developed a new method called H2E (Human-to-Embodied) to improve the performance of vision-language-action (VLA) models in embodied navigation tasks. Instead of relying on a large language model (LLM) planner for each navigation step, H2E distills successful agent experiences directly into the VLA model. This allows the VLA to navigate independently, reducing cost and delay. The H2E method converts complex agent sessions, including detours and sub-instructions, into training data for the VLA, focusing on successful goal achievement and masking irrelevant movements. This approach significantly boosts the success rate of an 8B VLA model on the REVERIE-CE benchmark, increasing it from 35.0% to 44.9%. AI
IMPACT This research could lead to more efficient and capable embodied AI agents, reducing reliance on costly LLM planning for navigation tasks.
RANK_REASON The item describes a new method and its performance on a benchmark, fitting the research category. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →