LUCID Enables Long-Horizon Humanoid Loco-Manipulation with Latent Skills

Achieving long-horizon loco-manipulation in humanoids requires combining versatile whole-body skills with effective high-level decision-making. Existing methods often rely on pre-trained skills coordinated by scripted planners or task-specific policies, which limits their adaptability to complex, sequential tasks.LUCID (Latent-Skill Unified Control via Imagined Dynamics) addresses this by proposing a hierarchical model-based reinforcement learning framework. It first trains a structured, latent-conditioned low-level policy using adversarial imitation, which is then frozen. Concurrently, a high-level policy and a macro-dynamics world model are learned.The world model predicts extended state transitions based on latent decisions, allowing the high-level policy to optimize through imagined rollouts. Evaluated in simulated multi-object rearrangement scenarios, LUCID demonstrates improved full-task success and partial-completion rates compared to previous baselines, showcasing its effectiveness for complex sequential tasks.