Overview
The research introduces PAMoR (Parameterized Affective Motion Generation in Real Time), a system designed for generating expressive whole-body motion in humanoid robots. This system addresses the challenge of creating robot movements that convey affect, which is crucial for social interactions. Unlike previous approaches that rely on reference clips or emotion words, PAMoR quantifies affect as a measurable control parameter: a valence-arousal (V-A) coordinate. This coordinate is derived directly from robot kinematics, specifically postural expansion and movement energy, eliminating the need for human annotation.
Research Context
Human perception of humanoid robot motion in social settings extends beyond the action performed to encompass the affect conveyed. Historically, generating motion with specific affective styles has involved methods such as extracting style from reference clips or associating it with emotion words. These methods, however, lack quantitative parameterization, limiting precise control and real-time modulation of affect during motion generation. The existing landscape for human avatars has predominantly used these less quantifiable methods for style and emotion.
Approach
PAMoR operates by computing a valence-arousal (V-A) coordinate natively on robot kinematics. This computation is performed in closed form, utilizing measurements of postural expansion and movement energy. These derived measurements directly serve as the conditions for motion generation, bypassing the requirement for human annotation. The system utilizes a shared latent space for training an action prior and two distinct affect priors. During each denoising step, these priors are composed: the action prior dictates 'what' is performed, while the two affect priors modulate 'how' it is performed. Motion generation is autoregressive, producing whole-body movements for a 29-DoF Unitree G1 robot in real time. Both the action and affect components of the generated motion are editable during operation.
Findings
- PAMoR successfully generates whole-body motion autoregressively on a 29-DoF Unitree G1 robot in real time.
- The system computes a valence-arousal (V-A) coordinate from postural expansion and movement energy, using these measurements as generation conditions without human annotation.
- Generated motion tracks the commanded V-A across its full range.
- The fidelity of text-to-motion generation by PAMoR matches that of text-only baselines.
- In a perceptual study, human raters identified the commanded emotion in 0.38 of trials. This performance exceeds both baselines used in the study.
- The identification rate of 0.38 approaches the 0.44 rate reported for acted human bodies.