LiMoDE: Mixture-of-Experts Lets Robots Learn Forever Without Forgetting
A dynamic expert system helps robots continuously adapt without catastrophic forgetting.
LiMoDE (Lifelong Mixture of Dynamic Experts) tackles catastrophic forgetting in robotics by rethinking architecture. In stage one (multi-task pre-training), a dynamic MoE activates varied numbers of motion-based experts for short-horizon tasks, building reusable prior knowledge. In stage two (lifelong adaptation), new experts are learned per task and combined with frozen prior experts via a dynamic gating mechanism. This allows robots to continuously add skills without overwriting old ones. Evaluated on both simulated lifelong benchmarks and real-world robot manipulation tasks, LiMoDE outperforms prior fine-tuning and prompt-based methods while adding only moderate parameter and inference overhead. The paper (arXiv:2606.26183) shows strong lifelong adaptation, making generalist robots that learn across tasks a step closer to reality.
- Two-stage scheme: dynamic MoE pre-training then lifelong expert adaptation with frozen prior knowledge.
- Activates heterogeneous experts based on motion information for short-term manipulation tasks.
- Outperforms fine-tuning and prompt-based methods on simulated and real-world tests with minimal extra parameters.
Why It Matters
Enables robots to continuously learn new skills without losing old ones, moving toward generalist, lifelong-learning robots.