X-Planner is a planning front-end for Vision-Language-Action systems that makes long-horizon task planning explicit, combining teleoperation data at varying annotation granularity with takeover-time and human-designed failure annotations for error recognition. A shared VLM backbone exposes two event-structured plan representations, a discrete interpretable interface and a latent interface that relays continuous chain-of-thought states across staggered Transformer depths via Staircase Decoding, anchored by a frozen latent-to-text reconstruction objective. It ranks second of four models on offline planning-quality metrics and outperforms baselines in real-robot execution experiments.
