Latent Adaptive Planner for Dynamic Manipulation

Donghun Noh, Deqian Kong, Minglu Zhao, Andrew Lizarraga, Jianwen Xie, Ying Nian Wu, Dennis Hong·May 06, 2025

Summary

LAP, a nonprehensile manipulation technique, uses human demonstration videos for latent space inference. It employs a principled variational replanning framework for visuomotor policy learning, ensuring temporal consistency and efficient adaptation. Bayesian updating in the latent space boosts success rate, trajectory smoothness, and energy efficiency. LAP effectively bridges the embodiment gap, enabling complex, human-like adaptable interactions across various robotic platforms through shared human demonstration videos. An optimal replanning horizon of 10 enhances real-time performance.

Introduction
Background
Overview of nonprehensile manipulation
Importance of human demonstration videos in robotic learning
Objective
Aim of using LAP in robotic manipulation
Key benefits of the approach
Method
Data Collection
Source of human demonstration videos
Preprocessing steps for video data
Data Preprocessing
Techniques for preparing data for analysis
Importance of data quality in learning outcomes
Principled Variational Replanning Framework
Explanation of the framework
How it ensures temporal consistency and efficient adaptation
Bayesian Updating in the Latent Space
Mechanism of Bayesian updating
Impact on success rate, trajectory smoothness, and energy efficiency
Bridging the Embodiment Gap
Challenges in robotic manipulation
How LAP addresses these challenges
Optimal Replanning Horizon
Determination of the optimal replanning horizon
Real-time performance enhancement
Results
Success Rate Improvement
Quantitative analysis of success rate increase
Trajectory Smoothness
Evaluation of trajectory smoothness
Energy Efficiency
Assessment of energy consumption
Complex, Human-like Interactions
Demonstration of adaptable interactions across robotic platforms
Case Studies
Examples showcasing the effectiveness of LAP
Conclusion
Summary of Key Findings
Future Directions
Potential improvements and advancements
Implications for Robotics and AI
Broader impact on robotic manipulation and learning
Basic info
papers
robotics
machine learning
artificial intelligence
Advanced features