MaMoRetarget: Morphology-Aware Human-to-Robot Motion Retargeting Across Humanoid Embodiments

Abstract

Human-to-robot motion retargeting can be viewed as a structured cross-modal motion translation problem, where the semantic intent and temporal dynamics of human motion must be preserved, while respecting embodiment-specific topology, joint constraints, and physical geometry. Existing optimization-based methods can provide geometrically plausible initial mappings, but they are typically per-sequence and non-amortized, and often suffer from self-collisions, unsafe near-contacts, and local temporal artifacts. In this paper, we propose a two-stage framework for motion retargeting across diverse humanoid embodiments. In the first stage, we introduce a minimum-change refinement procedure that repairs collision-prone local kinematic chains through discrete local search. This produces physically improved supervisory trajectories while preserving motion semantics and temporal continuity. We then propose MaMoRetarget, a morphology-aware neural retargeting network that learns an end-to-end feed-forward mapping from structured human motion and robot specifications to executable robot motion. MaMoRetarget integrates a morphology-conditioned human motion encoder, a kinematic tree-based robot specification encoder, and a cross-attention decoder with explicit joint-limit handling, enabling joint modeling of human dynamics and robot-specific embodiment constraints. Experiments on a MoCap-based multi-robot benchmark show that our framework improves physical plausibility and temporal smoothness, supports downstream policy learning, and generalizes well across diverse humanoid embodiments. These results demonstrate that MaMoRetarget offers an efficient and scalable paradigm for high-quality humanoid motion retargeting.