The predecessor to this work, hdmove1 , established a baseline for using low-dimensional latent embeddings for motion planning but suffered from projection instability when reconstructing high-frequency movements [3]. This paper presents , which overcomes these limitations through three primary contributions: