AdaMorph: Unified Motion Retargeting via Embodiment-Aware Adaptive Transformers

TL;DR

AdaMorph tackles cross-robot motion retargeting by learning a morphology-agnostic latent intent space and conditioning generation with a dual-pathway prompting mechanism and AdaLN. A canonical base-frame representation, differentiable integration, and gravity projection enforce physical plausibility, enabling a single model to control diverse humanoids without embodiment-specific retraining. Empirical results across 12 robots show high semantic preservation, strong rhythmic alignment, and zero-shot generalization to unseen motions, demonstrating robust cross-embodiment transfer. The approach reduces per-robot engineering while maintaining dynamic fidelity, with potential for real-time sim-to-real deployment and physics-informed extensions.

Abstract

Retargeting human motion to heterogeneous robots is a fundamental challenge in robotics, primarily due to the severe kinematic and dynamic discrepancies between varying embodiments. Existing solutions typically resort to training embodiment-specific models, which scales poorly and fails to exploit shared motion semantics. To address this, we present AdaMorph, a unified neural retargeting framework that enables a single model to adapt human motion to diverse robot morphologies. Our approach treats retargeting as a conditional generation task. We map human motion into a morphology-agnostic latent intent space and utilize a dual-purpose prompting mechanism to condition the generation. Instead of simple input concatenation, we leverage Adaptive Layer Normalization (AdaLN) to dynamically modulate the decoder's feature space based on embodiment constraints. Furthermore, we enforce physical plausibility through a curriculum-based training objective that ensures orientation and trajectory consistency via integration. Experimental results on 12 distinct humanoid robots demonstrate that AdaMorph effectively unifies control across heterogeneous topologies, exhibiting strong zero-shot generalization to unseen complex motions while preserving the dynamic essence of the source behaviors.

Figures (5)

• Figure 1: Overview of the proposed AdaMorph framework. The architecture unifies cross-embodiment retargeting through a Canonical Base-Frame Representation that standardizes human motion features into local velocities and articulation ($\mathbf{v}_t, \boldsymbol{\omega}_t, \mathbf{g}_t, \mathbf{J}_t$). (Left) A Morphology-Agnostic Intent Encoder maps these inputs, conditioned on SMPL shape parameters $\boldsymbol{\beta}$ via a Dynamic Human Prompt, to a shared latent intent $\mathbf{z}_t$. (Right) To bridge kinematic disparities, Learnable Static Robot Prompts ($\mathbf{P}_r$) drive the decoding process using AdaLin-Style Modulation, injecting embodiment-specific priors into the shared stream before Embodiment-Specific Output Adapters project the motion onto the target robot's joint space.
• Figure 2: Qualitative validation in the MuJoCo physics simulator. The unified model successfully retargets input human motions to all 12 trained robot embodiments. Despite the differences in link lengths and joint configurations, the robots faithfully reproduce the source behaviors, demonstrating the efficacy of our soft-prompted unified architecture.
• Figure 3: Visualization of Learned Robot Representations. (a) Cosine Similarity Matrix: The block-diagonal structure indicates high correlation between robots with similar kinematic chains (e.g., the Unitree family), proving that the model captures topological similarities. (b) t-SNE Projection: The projection of 16 learnable tokens per robot reveals semantic clustering. The model automatically groups robots by morphological similarity (e.g., G1 and H1 are proximal), forming stable identity signatures.
• Figure 4: Quantitative Evaluation of Semantic Consistency. (a) Root Velocity Consistency: The PCC between the root speed profiles of the human input and robot output. High correlations across diverse embodiments indicate that the robots accurately follow the human's movement rhythm (e.g., acceleration and deceleration). (b) Whole-Body Activity Consistency: The PCC of the mean joint velocity magnitudes. The consistently high scores demonstrate that the model effectively transfers the overall energy and intensity of the motion, regardless of the robot's specific kinematic configuration.
• Figure 5: Zero-Shot Retargeting on Unseen Folk Dances. Snapshots of diverse robots performing complex ethnic dance movements. Although these specific motion styles and subjects were completely absent from the training data, the unified model successfully transfers the intricate footwork and posture to the robot embodiments without any fine-tuning.