Reproduce “LUCID: Latent-Skill Unified Control via Imagined Dynamics for Long-Horizon Humanoid Loco-Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cheng Guo, Mingzhe Ni, Angelo Cangelosi, Arash Ajoudani
LUCID (Latent-Skill Unified Control via Imagined Dynamics) is a hierarchical model-based reinforcement learning framework that enables humanoids to perform complex, long-horizon loco-manipulation tasks like multi-object rearrangement. It achieves this by composing versatile whole-body skills and making high-level decisions via imagined macro-transitions, demonstrating improved task completion and generalization over baselines.
Key insights
- Planning over task-level macro-transitions via an imagined dynamics model is more effective and stable for long-horizon humanoid control than predicting fine-grained joint-level dynamics.
- A structured latent space for low-level skills, combining discrete semantic anchors with continuous variations, is crucial for high-level policies to make interpretable and task-aligned decisions.
- Jointly training a predictive world model with a high-level policy allows for foresightful optimization of skill sequences, leading to robust coordination of complex, sequential loco-manipulation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.