DeepTransition releases embodied foundation model DELE-w0.5 with a new world model approach

DELE-w0.5, a world-action model from DeepTransition, was released on September 4. During training it jointly predicts actions and the future world representation after the action; at inference it generates actions directly from language and current vision, without producing future video. The company reports 62.5% full-task success across 640 real robot trials and median inference latency of 87.5 milliseconds. DeepTransition previously raised an angel round from Fuxing Chuangfu as sole investor.

Original sources (Chinese)