Sample-efficient Low-level Motion Planning for Robotic Manipulation Tasks via Zero-shot Transfer Learning

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

The iCEM+TL framework introduces a novel approach to sample-efficient low-level motion planning for complex robotic manipulation tasks, addressing challenges like long training times and limited performance of existing evolutionary algorithms. This framework integrates Transfer Learning (TL) and Reward Redesign (RR) into the Sample-efficient Cross-Entropy Method (iCEM). TL involves transferring key iCEM parameters and elite trajectories from simpler upstream tasks to guide more complex downstream tasks like stacking, sliding, and shelf placement in a zero-shot manner. RR optimizes task-specific performance through reward decomposition for stacking and shelf placement. Simulation results demonstrate success rate improvements of up to 23% over baselines on FetchStack, FetchSlide, and a new Shelf environment. The framework's practical feasibility was further validated on a real Franka Emika robot in a stacking task.

Key takeaway

For robotics engineers developing manipulation systems, the iCEM+TL framework offers a computationally efficient alternative to purely learning-based approaches. If you are facing challenges with long training times or limited performance in complex tasks like stacking or shelf placement, consider implementing zero-shot transfer learning and reward redesign. This can significantly improve success rates, as demonstrated by up to 29.3% gains on the Stack task, without extensive retraining.

Key insights

The iCEM+TL framework enhances robotic motion planning by combining zero-shot transfer learning and reward redesign for complex manipulation tasks.

Principles

• Transfer learning benefits from structural alignment.
• Task decomposition improves scalability.
• TL and RR combine for significant performance gains.

Method

The iCEM+TL framework transfers Gaussian sampling distribution parameters (μ, σ) and elite trajectories from upstream tasks, then applies Reward Redesign by decomposing complex objectives into R_base, R_lift, and R_gripper for downstream tasks.

In practice

• Use zero-shot transfer for complex manipulation tasks.
• Decompose rewards for stacking and shelf placement.
• Select upstream tasks with positive elite reward advantage.

Topics

• Robotic Manipulation
• Motion Planning
• Transfer Learning
• Evolutionary Algorithms
• Reward Redesign
• iCEM Framework

Code references

• YuanzhiHe/iCEM_TL

Best for: Research Scientist, AI Scientist, Robotics Engineer

Related on AIssential

• Sample-efficient Low-level Motion Planning for Robotic Manipulation Tasks via Zero-shot Transfer Learning — Leveraging transfer learning and reward redesign significantly improves robotic motion planning efficiency.
• Jump-Start Reinforcement Learning with Vision-Language-Action Regularization — VLAJS jump-starts RL with sparse, transient VLA guidance, improving sample efficiency and real-world robot control.
• ASPECT:Analogical Semantic Policy Execution via Language Conditioned Transfer — LLMs can dynamically remap task descriptions for zero-shot transfer in RL agents.
• From Kinematics to Dynamics: Learning to Refine Hybrid Plans for Physically Feasible Execution — Robotic plans based on first-order dynamics are physically infeasible; a neuro-symbolic RL framework can refine them for second-order feasibility.
• Gradient-based planning for world models at longer horizons — GRASP enables robust long-horizon planning for world models by parallelizing optimization and improving gradient signals.
• TRN-R1-Zero: Text-rich Network Reasoning via LLMs with Reinforcement Learning Only — TRN-R1-Zero enables zero-shot reasoning on text-rich networks using reinforcement learning without supervised fine-tuning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.