ESCAPE: Episodic Spatial Memory and Adaptive Execution Policy for Long-Horizon Mobile Manipulation

2026-04-15

AI Summary

Long-horizon mobile manipulation faces significant challenges such as catastrophic forgetting, spatial inconsistency, and execution rigidity, which hinder effective coordination between navigation and manipulation. This work proposes a tightly coupled perception–localization–execution framework that leverages a depth-agnostic, episodic 3D spatial memory mechanism and an adaptive execution policy to dynamically coordinate global navigation with local manipulation. The approach enables opportunistic target acquisition and efficient exploration even without detailed instructions. Evaluated on the ALFRED benchmark, the method achieves state-of-the-art performance, with task success rates of 65.09% and 60.79% in seen and unseen environments, respectively, along with substantial improvements in path-weighted metrics. Notably, it maintains robust performance under instruction-free conditions, achieving success rates of 61.24% and 56.04% in seen and unseen settings.

Technology Category

• Intelligent Robots: Manipulation
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Coordinating navigation and manipulation with robust performance is essential for embodied AI in complex indoor environments. However, as tasks extend over long horizons, existing methods often struggle due to catastrophic forgetting, spatial inconsistency, and rigid execution. To address these issues, we propose ESCAPE (Episodic Spatial Memory Coupled with an Adaptive Policy for Execution), operating through a tightly coupled perception-grounding-execution workflow. For robust perception, ESCAPE features a Spatio-Temporal Fusion Mapping module to autoregressively construct a depth-free, persistent 3D spatial memory, alongside a Memory-Driven Target Grounding module for precise interaction mask generation. To achieve flexible action, our Adaptive Execution Policy dynamically orchestrates proactive global navigation and reactive local manipulation to seize opportunistic targets. ESCAPE achieves state-of-the-art performance on the ALFRED benchmark, reaching 65.09% and 60.79% success rates in test seen and unseen environments with step-by-step instructions. By reducing redundant exploration, our ESCAPE attains substantial improvements in path-length-weighted metrics and maintains robust performance (61.24% / 56.04%) even without detailed guidance for long-horizon tasks.

Problem

• long-horizon mobile manipulation
• catastrophic forgetting
• spatial inconsistency
• rigid execution
• embodied AI

Innovation

• Episodic Spatial Memory
• Adaptive Execution Policy
• Spatio-Temporal Fusion Mapping
• Memory-Driven Target Grounding
• Long-Horizon Mobile Manipulation