LUCID: Hierarchical framework that unifies skills with imagined dynamics for long-term humanoid loco-manipulation

Summary

LUCID is a hierarchical, model-based reinforcement learning framework. It trains a low-level whole-body skill policy via adversarial imitation and then freezes it. Subsequently, it learns a high-level policy and a macro-dynamic world model to plan these skills through imagined simulations. The goal is long-horizon humanoid locomotion and manipulation tasks.

Why it matters

Combining a learned world model with reusable whole-body skills is a natural path toward general-purpose humanoid manipulation. A clear hierarchical decomposition (frozen low-level + learned high-level) can reduce the combinatorial burden of end-to-end policies. If imagined simulation planning extends beyond simulation, it could lower the data and computational costs that currently pose challenges in non-model-based humanoid control.

Highlights

• The low-level policy is trained once via adversarial imitation and then frozen. This decouples skill acquisition from task planning.
• The macro-dynamic world model predicts state transitions over time. This allows the high-level policy to optimize over imagined simulations rather than real ones.
• Evaluated on simulated multi-object rearrangement tasks. Improvements in full task success and partial completion rates were reported compared to scripted planners and non-model-based baselines.

Critical take

All results are limited to simulated multi-object rearrangement tasks. The compared baselines (finite state machines, scripted planners, task-specific non-model-based policies) are relatively weak. This makes it difficult to contextualize the claimed improvements when compared to state-of-the-art model-based agents like Dreamer-style architectures. There is no discussion regarding real-robot validation, ablation studies on world model fidelity, computational load, or generalization to unseen objects. These gaps leave questions about practical deployability.