Continuous affect-conditioned whole-body humanoid motion generation

Establish continuous affect-conditioned whole-body motion generation for humanoid robots, extending affective motion control beyond discrete labels, reference examples, and limited robot platforms.

Background

The paper identifies a gap between affective motion generation for human avatars and real-time generation for fully articulated humanoid robots. Existing avatar systems specify affect through reference clips or emotion labels, while robot systems use offline optimization, latent-space editing, or short expressive animations. The authors state that conditioning whole-body humanoid generation directly on a continuous valence–arousal signal remains unresolved; PAMoR is presented as an effort to address this gap on the 29-DoF Unitree G1.

References

Affective and expressive motion has been generated in several ways, but conditioning whole-body humanoid generation on a continuous affect signal remains open.

— PAMoR: Parameterized Affective Motion Generation in Real Time for Humanoid Robots  (2608.28213 - Pan et al., 28 Aug 2026) in Section 2.3, “Affective Motion Generation”

Whether closing this loop holds the commanded affect on the real robot is the question we would take up first.

— PAMoR: Parameterized Affective Motion Generation in Real Time for Humanoid Robots  (2608.28213 - Pan et al., 28 Aug 2026) in Section 6, “Discussion”