Skip to main content
Back

iCEM+TL Lets Robots Stack and Slide Objects Without Retraining From Scratch

Yuanzhi He et al. propose iCEM+TL, a zero-shot transfer learning framework that boosts robotic manipulation success rates by up to 23% with no retraining required.

  • iCEM+TL framework: Yuanzhi He et al. propose transferring iCEM planner parameters zero-shot from simple to complex robotic tasks, accepted to ICANN 2026. [Paper]
  • Up to 23% success rate improvement over baseline iCEM across stacking, sliding, and shelf placement tasks in MuJoCo simulation.
  • Zero-shot transfer: Key optimizer hyperparameters (noise exponent, elite-set size, planning horizon) are reused from solved upstream tasks with no retraining.
  • Reward Redesign (RR) decomposes complex tasks like stacking into sub-goals, providing denser reward signals to the planner.
  • Real-robot validation on a Franka Emika FR3 confirms sim-to-real feasibility for the stacking task using MuJoCo-optimized trajectories.
  • No public code released; sim-to-real transfer is currently open-loop, limiting use in unstructured or dynamic environments.

Training a robot to stack blocks is hard. Training it to also slide objects, place items on shelves, and generalize across all of these tasks without starting from scratch each time is even harder. A new research paper by Yuanzhi He and colleagues, accepted to ICANN 2026, proposes a framework called iCEM+TL that tackles exactly this problem, combining an established evolutionary planner with zero-shot transfer learning to let robots reuse knowledge from simpler tasks when tackling more complex ones.

The wall that iCEM hits

Evolutionary algorithms such as the Sample-efficient Cross-Entropy Method (iCEM) have recently demonstrated promising potential for low-level real-time planning by leveraging efficient knowledge reuse strategies to improve performance. A quick primer: iCEM is a trajectory optimizer that works inside a Model Predictive Control (MPC) loop. Rather than learning a neural policy end-to-end, it repeatedly samples candidate action sequences, scores them against a reward function, keeps the best ones (the "elite set"), and refines its sampling distribution around those winners. The original iCEM requires 2.7 to 22 times fewer samples than standard CEM and yields a performance increase of 1.2 to 10 times in high-dimensional control problems.

Although effective in many control tasks, iCEM's performance can be constrained in more complex scenarios, particularly those requiring stacking, sliding, and shelf placement. The root cause is that iCEM's internal parameters, things like the noise exponent that shapes how correlated action sequences are, the elite-set size, and the planning horizon, are tuned per task. When complexity increases, those defaults no longer hold, and performance degrades without a principled way to bootstrap from what the planner already knows.