Reproduce “Mobile Manipulation with Active Inference for Long-Horizon Rearrangement Tasks” on OpenResearch

About this paper

Corrado Pezzato, Ozan Çatal, Toon Van de Maele, Riddhi J. Pitliya, Tim Verbelen

This paper from VERSES AI Research Lab presents a fully hierarchical active inference system capable of performing long-horizon mobile manipulation tasks. The system achieves higher success rates on the Habitat Benchmark compared to state-of-the-art reinforcement learning baselines, notably without requiring extensive offline training data.

Key insights

• Active inference, typically applied in simplified settings, can scale to complex, long-horizon mobile manipulation tasks by leveraging a fully hierarchical architecture.
• A unified, online inference-driven approach, integrating high-level discrete planning with low-level continuous control, can outperform data-intensive, offline-trained reinforcement learning methods.
• The inherent properties of active inference, such as continuous belief updating and explicit retry mechanisms, enable strong generalization and robust online adaptation to task failures without extensive prior training.

Summary by alphaXiv.