iCEM+TL boosts robot motion planning by 23% with zero-shot transfer learning

Robotic motion planning is getting a major efficiency upgrade from a new framework called iCEM+TL, developed by a team led by Yuanzhi He. The approach enhances the existing iCEM (sample-efficient Cross-Entropy Method) — an evolutionary algorithm used for low-level real-time planning — by explicitly incorporating transfer learning (TL). Instead of training from scratch for every complex manipulation task (e.g., stacking objects, sliding, or shelf placement), the system reuses key iCEM parameters learned from simpler upstream tasks. This zero-shot transfer dramatically cuts training time and improves performance in more challenging scenarios.

The researchers further optimized performance through Reward Redesign (RR), which decomposes tasks like stacking into sub-goals for more targeted reward signals. In simulation, iCEM+TL achieved success rate improvements of up to 23% over standard iCEM. The team validated the framework on a real Franka Emika robot arm performing a stacking task, confirming its practical feasibility for real-world deployment. The paper has been accepted at the International Conference on Artificial Neural Networks (ICANN) 2026.