Multi-Expert Self-Improving HY-Embodied: Combining Hierarchical MoE and Online Skill Acquisition

TL;DR: Can HY-Embodied-0.5 keep learning new skills on its own, even after deployment? By fusing Hierarchical Mixture-of-Experts (HiMoE-VLA, Du et al., 2025) with self-improvement post-training (Seyed Ghasemipour et al., 2025), we’ll enable continual autonomous skill acquisition in the wild.

Research Question: How effective is a self-improving, hierarchical Mixture-of-Experts architecture in extending HY-Embodied-0.5’s capabilities through ongoing, autonomous practice and specialization?

Hypothesis: Integrating hierarchical expert modules with online self-improvement will yield a model that not only generalizes across heterogeneous tasks and robots but also continually enhances its own abilities with minimal human oversight.

Experiment Plan: - Implement HiMoE action modules within HY-Embodied-0.5 to allow specialization for different domains or robot types.

• Add a self-improvement loop: after initial deployment, allow the model to autonomously practice, extract reward signals, and refine policies.
• Deploy in a multi-task, multi-robot environment; track skill acquisition, specialization, and performance gains over time.
• Compare to static and non-hierarchical models in terms of adaptability, sample efficiency, and final task competence.

References:

• Du, Z., Liu, B., Liang, Y., Shen, Y., Cao, H., Zheng, X., Feng, Z., Wu, Z., Yang, J., & Jiang, Y.-G. (2025). HiMoE-VLA: Hierarchical Mixture-of-Experts for Generalist Vision-Language-Action Policies. arXiv.org.
• Seyed Ghasemipour, S. K., Wahid, A., Tompson, J., Sanketi, P. R., & Mordatch, I. (2025). Self-Improving Embodied Foundation Models. arXiv.org.
• X. TencentRobotics et al. (2026). HY-Embodied-0.5: Embodied Foundation Models for Real-World Agents.