Humanoid-DART: Humanoid Loco-Manipulation using Diffusion-guided Augmentation through Relabeling and Tracking

2026-06-25

AI Summary

This work addresses the limitations of existing approaches to humanoid robot loco-manipulation—namely, their heavy reliance on extensive human demonstrations, poor scalability, and need for continuous intervention—by introducing a self-supervised learning framework. Starting from sparse human demonstrations, the method leverages a diffusion model to generate goal-conditioned trajectories, which are then refined through reinforcement learning for accurate trajectory tracking. A trajectory relabeling mechanism is incorporated to enable autonomous exploration of the goal space and facilitate skill expansion. With minimal expert supervision, the proposed framework substantially enhances behavioral diversity, generalization capability, and sample efficiency, outperforming current state-of-the-art methods across a range of loco-manipulation tasks.

Abstract

Imitating human demonstrations has emerged as a dominant paradigm for learning humanoid loco-manipulation policies. However, scaling these approaches remains challenging due to the high cost of collecting diverse demonstrations and the need for continual human intervention to correct policy failures. In this paper, we present a self-supervised framework that bootstraps from sparse demonstrations and progressively expands its behavioral repertoire, enabling the learning of a goal-conditioned policy that automatically explores the goal space with minimal expert supervision. Our approach combines diffusion-based trajectory generation with reinforcement learning, where the latter is used to track goal-conditioned trajectories produced by the diffusion model for a range of loco-manipulation skills. Through extensive ablation studies and comparisons with state-of-the-art methods, we demonstrate the effectiveness of our framework on multiple humanoid loco-manipulation skills.

Problem

• humanoid locomotion
• manipulation
• imitation learning
• self-supervised learning
• goal-conditioned policy

Innovation

• diffusion-guided augmentation
• self-supervised learning
• goal-conditioned policy
• loco-manipulation
• humanoid robotics