X-Planner Open-Sourced: Event-Structured Task Planning with Staircase Latent Decoding for Embodied Manipulation

Core Background & Industry Pain Points Long-horizon manipulation (e.g., clearing cluttered countertops, nested assembly) requires coordinating high-level semantic intentions with low-level joint execution. Contemporary Vision-Language-Action (VLA) models obscure task reasoning inside black-box weights, rendering recovery impossible upon contact slips. Conversely, serializing explicit natural language Chain-of-Thought (CoT) word-by-word incurs severe latency penalties that cripple real-time reactive control. ### Architecture Highlights & Internals X-Planner introduces an event-structured planning front-end featuring: (1) Dual Interfaces on a shared VLM backbone, pairing a human-readable discrete event state output with an ultra-compact continuous latent representation; (2) Staircase Decoding, which pipelines continuous latent thoughts across staggered Transformer depths to bypass token generation latency; and (3) A frozen latent-to-text auxiliary reconstruction anchor that penalizes semantic drift in continuous representations. ### Authoritative Benchmarks & Measured Scores Evaluated on offline benchmarks and real-world robotic manipulators: X-Planner achieves an 82.3% real-robot task completion rate across multi-phase manipulation episodes (versus 54.1% for end-to-end VLAs and 61.8% for tokenized CoT baselines). Staircase Decoding slashes decision-making latency by 48.5%, achieving the response velocity necessary for reactive embodied deployment. ### Developer Hands-on Guide Robotics engineers can view hardware trials at https://x-square-robot.github.io/Xplanner/ and clone the open-source PyTorch ROS2 package from GitHub (X-Square-Robot/Xplanner).