Human-Humanoid Robots Cross-Embodiment Behavior-Skill Transfer Using Decomposed Adversarial Learning from Demonstration

TL;DR

The paper tackles the data bottleneck and cross-embodiment transfer challenge in humanoid loco-manipulation by introducing a Unified Digital Human (UDH) as a common prototype and decomposing high-DoF control into learned behavior primitives via Decomposed Adversarial Imitation Learning (DAIL). It couples kinematic motion retargeting with an interaction-graph guided high-level policy to plan latent behaviors that coordinate across body parts, followed by embodiment-specific fine-tuning with an MLP to map to dynamics. Key contributions include the UDH design with 92 DoFs, per-part adversarial learning for primitives, the interaction graph framework with a style discriminator, and demonstration across five diverse humanoids with improved data efficiency and transfer performance. The approach promises practical impact for rapid deployment of humanoid skills across platforms, reducing data needs while preserving natural, stable loco-manipulation behavior.

Abstract

Humanoid robots are envisioned as embodied intelligent agents capable of performing a wide range of human-level loco-manipulation tasks, particularly in scenarios requiring strenuous and repetitive labor. However, learning these skills is challenging due to the high degrees of freedom of humanoid robots, and collecting sufficient training data for humanoid is a laborious process. Given the rapid introduction of new humanoid platforms, a cross-embodiment framework that allows generalizable skill transfer is becoming increasingly critical. To address this, we propose a transferable framework that reduces the data bottleneck by using a unified digital human model as a common prototype and bypassing the need for re-training on every new robot platform. The model learns behavior primitives from human demonstrations through adversarial imitation, and the complex robot structures are decomposed into functional components, each trained independently and dynamically coordinated. Task generalization is achieved through a human-object interaction graph, and skills are transferred to different robots via embodiment-specific kinematic motion retargeting and dynamic fine-tuning. Our framework is validated on five humanoid robots with diverse configurations, demonstrating stable loco-manipulation and highlighting its effectiveness in reducing data requirements and increasing the efficiency of skill transfer across platforms.

Figures (7)

• Figure 1: Schematic overview of the cross-embodiment loco-manipulation skill transfer framework. 1) Human embodiment demonstration is captured by motion capture system and retargeted to the unified digital human, then retargeted to diverse humanoid robots. 2) High DoFs of humanoids are decomposed into functional parts $p\in [1, P]$ and trained with partial demonstration separately via adversarial imitation to form the latent behavior primitive spaces, humanoids can perform natural and coordinated motions by the behavior controller $\pi(\textbf{a} | \textbf{s}, \textbf{z}_{1\sim P})$. 3) Interaction graphs are extracted from demonstration and guides the policy $\eta( \textbf{z}_{1\sim P}|\textbf{s}, \mathcal{G}, g)$ learning to plan latent behavior trajectories on behavior spaces to complete the interaction skill. 4) By kinematic motion retargeting and adversarial imitation fine-tuning on specific embodiment, same loco-manipulation motions can be deployed on diverse humanoid robots.
• Figure 2: Kinematic motion retargeting. (a) Motions are retargeted by grouping DoFs belonging to same function parts and each joint value is solved by partial inverse kinematics. Orange ellipses refer to the group of shoulder DoFs, light blue ellipses refer to the group of hip DoFs. (b) The animation screenshots of kinematic motion retargeting on the unified digital human, NAVIAI and H1 humanoid robots. It is worth noting that the actual dynamics are not considered in this process (see the red boxes), and robots are merely simulating the joint angles, joint velocities, and the trajectories of root links.
• Figure 3: Whole-body functional decomposition and behavior primitive training. (a) Behavior primitive pre-training for the unified digital human and five humanoid robots. It shows the style imitation of the given motion dataset using pre-trained behavior primitives. (b) Hand-specific behavior primitive pre-training.
• Figure 4: Loco-manipulation skill learning with interaction graph. An illustration of the detailed loco-manipulation skill learning process with skill-level interaction graph guidance and decomposed behavior primitive trajectories in latent behavior spaces.
• Figure 5: Behavior primitive pre-training. Stability experiments for legged humanoid robots by hitting them with moving cubes.