MotionVLA: Vision-Language-Action Model for Humanoid Motion

A research team has introduced MotionVLA, a vision-language-action model that generates humanoid motion from scene images and text instructions, using a dual-stream frequency tokenizer to separate low-frequency pose data from high-frequency physical dynamics [1]. The model, detailed in a paper submitted to arXiv on June 13, 2026, is built on a Qwen3.5 backbone with 2 billion parameters [1][2]. Vision-language-action models, or VLAs, are a class of multimodal foundation models that integrate vision, language, and action outputs. In robot learning, a VLA takes an image of the surroundings and a text instruction, then directly outputs low-level actions executable by a robot [3]. The concept was pioneered by Google DeepMind with RT-2 in July 2023 [3]. More recently, robotics firm Figure AI has developed its own VLA, called Helix, to control its humanoid robots [4]. The MotionVLA paper addresses a specific technical bottleneck in motion generation. A frequency-domain analysis of human motion data found that five discrete cosine transform coefficients capture 93% of joint-position energy but only 37% of joint-velocity energy [1][2]. This mismatch means that single-codebook tokenizers, which force heterogeneous motion signals into one quantization space, tend to bias representation toward pose statistics and under-represent high-frequency velocity components [2]. To solve this, the authors propose a dual-stream frequency tokenizer, or DSFT, that splits motion into a Base stream and a physical stream, compressing each independently with DCT truncation and byte-pair encoding [1][2]. In the MotionVLA architecture, Base and physical tokens are arranged in a unified sequence, with physical tokens predicted after Base tokens [2]. The researchers tested the model on the HumanML3D and MBench datasets. On HumanML3D, MotionVLA reduced the Diversity gap to real data by over 50% [1][2]. On MBench, it improved Motion-Condition Consistency by 3.8% [1][2]. The paper’s authors argue that frequency-aware dual-stream decoupling is an effective formulation for autoregressive motion generation [2]. Code and a project website have been made publicly available [1]. arXiv, where the paper was posted, is an open-access repository of electronic preprints that are moderated but not peer-reviewed. It hosts papers in fields including computer science, physics, and mathematics, and as of November 2024 receives about 24,000 submissions per month [8].