ActionCodec advances VLA models with optimized action tokenization

Researchers have introduced ActionCodec, a novel approach to action tokenization for Vision-Language-Action (VLA) models. This method focuses on optimizing VLA performance rather than just reconstruction fidelity, establishing design principles such as maximizing temporal token overlap and minimizing vocabulary redundancy. When applied to a SmolVLM2-2.2B model, ActionCodec achieved a 95.5% success rate on the LIBERO benchmark without prior robotics training, setting a new state-of-the-art for VLA models. AI

IMPACT ActionCodec's principles could lead to more efficient and effective VLA models, accelerating progress in robotics and embodied AI.

RANK_REASON The cluster describes a new research paper detailing a novel method for action tokenization in AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ActionCodec
• arXiv
• Hugging Face
• LIBERO
• SmolVLM2-2.2B
• Vision-Language-Action (VLA)
• Vision Language Models (VLMs)
• Zibin Dong

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →