ActionCodec: What Makes for Good Action Tokenizers

This research paper presents ActionCodec, a high-performance action tokenizer that enhances training efficiency and VLA performance. The authors establish design principles based on information-theoret, including maximized temporal token overlap, minimized vocabulary redundancy, enhanced multimodal mutual information, and token independence. ActionCodec achieves a 95.5% success rate on LIBERO without robotics pre-training, with further improvements reaching 97.4% through architectural enhancements, setting a new SOTA for VLA models without pre-training.