OmniContact framework lets humanoids chain complex loco-manipulation at 98.7% success

Humanoid robots face a dual challenge: executing robust locomotion and manipulation (loco-manipulation) over long horizons while seamlessly chaining multiple skills with autonomous recovery. Existing methods either rely on explicit object-interaction representations that are precise but hard to plan with, or on implicit skill embeddings that lack interpretability. OmniContact bridges this gap with contact flow (CF), a compact representation of key body trajectories and binary contact signals over time. The framework consists of a low-level policy called CF-Track, which learns a unified library of loco-manipulation skills, and a high-level module CF-Gen that heuristically generates future contact-flow sequences.

In experiments, OmniContact achieves 98.7% success on the Carry Box task and 76.5% on the more complex Push-Stack Boxes task, outperforming prior baselines by average margins of 40.9% in meta-skill execution and 66.5% in skill chaining. The framework also naturally integrates with vision-language models (VLMs) for semantic task decomposition, enabling complex behaviors like arranging scattered boxes into a heart shape. The accompanying OmniContact dataset, a MoCap-based human-object interaction corpus, supports further research. This work marks a significant step toward generalizable humanoid manipulation in unstructured environments.