Beyond Dual-Streams: Integrating Homeostatic Motivation into XSkill for Emergent Skill Discovery

TL;DR: What if we taught XSkill agents to learn from their own internal "motivations" like hunger or curiosity, just like animals, to see if they can discover new skills on their own? We could augment XSkill with a homeostatic drive (e.g., energy, comfort), then let it operate in open-ended simulated environments to observe whether this internal feedback prompts the emergence of novel, unanticipated skills beyond those encoded via experience and skill streams.

Research Question: Can introducing homeostatic motivation into XSkill’s continual learning framework foster emergent skill discovery and more adaptive behavior in open-ended environments?

Hypothesis: Agents equipped with a homeostasis-driven intrinsic motivation, in addition to XSkill’s experience and skill streams, will autonomously acquire a broader and more integrated set of behaviors, surpassing the diversity and adaptability of skills learned via external task supervision alone.

Experiment Plan: - Setup: Extend XSkill’s architecture with a homeostatic module inspired by Yoshida & Kuniyoshi (2025), where internal states (e.g., "energy" or "curiosity") are tracked and must be kept within optimal ranges.

• Environment: Use rich, open-ended RL environments (e.g., Crafter or MiniGrid variants) with multiple affordances and survival challenges.
• Comparisons: Compare baseline XSkill, XSkill+homeostasis, and homeostasis-only (HRRL) agents in terms of skill diversity, task coverage, and adaptability to new challenges.
• Measurements: Quantify the emergence of novel skills, behavioral diversity, and robustness to environmental changes.
• Expected Outcome: XSkill+homeostasis agents will exhibit more varied and adaptive behaviors, including skills not prompted by explicit tasks, demonstrating the value of internal motivational feedback.

References:

• Jiang, G., Su, Z., Qu, X., & Fung, Y. R. (2026). XSkill: Continual Learning from Experience and Skills in Multimodal Agents.
• Yoshida, N., & Kuniyoshi, Y. (2025). Unexpected Capability of Homeostasis for Open-ended Learning. International Conference on Development and Learning.