Research Analyzer
← Back ICRA 2026

Not Throwing Away My Shot: Planning Ahead with Dual Subgoals in Long-Horizon Robot Manipulation Tasks

Longrui Chen, Yanlong Huang, Mehmet R Dogar

PDF

AI summary

Key figure (auto-extracted from paper)
Combining short-horizon and low-variance subgoals in a dual-subgoal framework significantly improves success rates and stability in long-horizon robotic manipulation tasks.
Long-horizon manipulation subgoal-conditioned learning dual subgoals imitation learning robotic planning hierarchical control

Problem

Subgoal-conditioned policies struggle with long-horizon tasks because identifying informative, consistent, and reachable intermediate subgoals is difficult, leading to high prediction variance and temporal fluctuation.

Approach

The authors propose PDS (Planning with Dual Subgoals), which jointly predicts and conditions on both a short-horizon subgoal and a low-variance subgoal within a shared embedding space using a conditional variational autoencoder and an RNN-based action policy.

Key results

  • Analyzed the influence of subgoal horizon and variance on hierarchical policy performance
  • Benchmarked time-, visual-, and language-based subgoal policies across simulation and real-world tasks
  • Introduced PDS framework combining short-horizon and low-variance subgoals to outperform single-subgoal baselines

Why it matters

Provides a practical framework for improving the reliability and success of long-horizon robotic manipulation, benefiting autonomous systems researchers and practitioners.

Abstract

Policy learning often encounters difficulties in long-horizon tasks. Subgoal-conditioned policies address long- horizon problems by decomposing them into manageable seg- ments, but they usually struggle with identifying informative subgoals. To address this limitation, we propose PDS (planning with dual subgoal), an architecture that learns short-horizon and low-variance subgoals in embedding space, ensuring the planning both reachable and consistent. We begin by analyzing the impact of horizon and consistency on the performance of subgoal-conditioned policies. We evaluate the performance of commonly used subgoal definitions (time-based, visual-based, and language-based) in tasks with different lengths. Subse- quently, we demonstrate that our approach, which predicts and conditions on dual subgoals, improves success rates and enhances stability across diverse tasks in simulation and real- world.

Index terms

Imitation Learning Manipulation Planning Deep Learning in Grasping and Manipulation

Related papers