A Hierarchical Approach to Imitation Learning for Manipulation Tasks Requiring Time Varying Forces

2026-08-04

AI Summary

This work addresses the challenge of applying diffusion-based policies to high-frequency force-controlled tasks, where inference latency hinders effective handling of contact-intensive disassembly operations—particularly transient fracture events. To overcome this limitation, the authors propose DPA-FTG, a hierarchical framework that decouples diffusion-based policy planning from real-time force regulation for the first time. At the high level, a conditional diffusion model operating at 5 Hz selects strategies from a task-primitive vocabulary; at the low level, a lightweight neural impedance controller runs at 60 Hz to achieve real-time visuo-tactile impedance control. This architecture balances expressive policy representation with dynamic contact stability, surpassing the open-loop constraints of existing action-chunking methods. Evaluated on a dual-arm robotic flexible battery disassembly task, DPA-FTG significantly outperforms baselines such as Reactive Diffusion Policy, achieving efficient and stable dynamic disassembly.

Abstract

Diffusion policies have shown strong performance in learning complex, multi-modal behaviors for robotic manipulation. However, their application to contact-rich disassembly tasks remains limited by a key trade-off: the iterative denoising process introduces inference latencies that makes high frequency control difficult, which is essential for realizing dynamic interactions such as chiseling and prying. Recent action-chunking techniques mitigate latency but use an open-loop execution window, rendering the system blind to rapid force transients caused by fracture events. To bridge this gap, we introduce the Diffusion Policy Augmented by Fast Trajectory Generation (DPA-FTG). Compared to recent visual-tactile approaches that focus on positional correction, DPA-FTG decouples low-frequency planning from high-frequency force regulation. At the high level ($5$ Hz), a conditional diffusion model predicts a sequence of latent parameters for selecting a strategy from a learned vocabulary of task primitives. At the low level ($60$ Hz), a lightweight, force-conditioned policy acts as a neural impedance controller, modulating execution in real-time to maintain contact stability. We validate our approach on a bimanual battery disassembly task involving the separation of a compliant sheet. Experimental evaluation demonstrates that DPA-FTG outperforms state-of-the-art baselines, including Reactive Diffusion Policy (RDP).

Problem

• imitation learning
• diffusion policy
• force control
• contact-rich manipulation
• disassembly tasks

Innovation

• Diffusion Policy
• Force Control
• Hierarchical Imitation Learning
• Neural Impedance Control
• Action Chunking