Learning Attentive Neural Processes for Planning with Pushing Actions

2025-04-24

AI Summary

This paper addresses tabletop pushing planning for objects with unknown physical properties. We propose an online autonomous learning framework based on partially observable Markov decision processes (POMDPs). Our key contributions are twofold: First, we introduce Attentive Neural Processes (ANPs) into POMDP state inference—replacing conventional particle filters—to enable efficient, differentiable, and online dynamics estimation. Second, we design the NPT-DPW planner, integrating Neural Process Trees with Double Progressive Widening to support end-to-end differentiable and scalable pushing-sequence generation. In simulation, our method achieves a 3.2× speedup in planning latency and a 27% improvement in task success rate over particle-filter-based baselines. Moreover, it demonstrates strong robustness under highly nonlinear dynamics and partial observability.

Technology Category

• Planning, Routing, and Scheduling: Planning with Markov Models (MDPs, POMDPs)
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Intelligent Robots: Motion and Path Planning

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Our goal is to enable robots to plan sequences of tabletop actions to push a block with unknown physical properties to a desired goal pose on the table. We approach this problem by learning the constituent models of a Partially-Observable Markov Decision Process (POMDP), where the robot can observe the outcome of a push, but the physical properties of the block that govern the dynamics remain unknown. The pushing problem is a difficult POMDP to solve due to the challenge of state estimation. The physical properties have a nonlinear relationship with the outcomes, requiring computationally expensive methods, such as particle filters, to represent beliefs. Leveraging the Attentive Neural Process architecture, we propose to replace the particle filter with a neural network that learns the inference computation over the physical properties given a history of actions. This Neural Process is integrated into planning as the Neural Process Tree with Double Progressive Widening (NPT-DPW). Simulation results indicate that NPT-DPW generates more effective plans faster than traditional particle filter methods, even in complex pushing scenarios.

Problem

• Enable robots to plan pushing actions for unknown block properties
• Replace particle filters with neural networks for state estimation
• Improve planning efficiency in complex pushing scenarios

Innovation

• Uses Attentive Neural Process for inference
• Replaces particle filter with neural network
• Integrates Neural Process Tree with DPW