LaGarNet: Goal-Conditioned Recurrent State-Space Models for Pick-and-Place Garment Flattening

TL;DR

LaGarNet is a goal-conditioned recurrent state-space model that effectively learns garment manipulation dynamics, achieving state-of-the-art flattening performance across various garment types in real and simulated environments.

Contribution

It introduces LaGarNet, the first state-space model for complex garment manipulation, reducing biases and matching mesh-based methods in performance.

Findings

Achieves state-of-the-art garment flattening results

Works effectively across multiple garment types

Operates successfully in both real-world and simulation environments

Abstract

We present a novel goal-conditioned recurrent state space (GC-RSSM) model capable of learning latent dynamics of pick-and-place garment manipulation. Our proposed method LaGarNet matches the state-of-the-art performance of mesh-based methods, marking the first successful application of state-space models on complex garments. LaGarNet trains on a coverage-alignment reward and a dataset collected through a general procedure supported by a random policy and a diffusion policy learned from few human demonstrations; it substantially reduces the inductive biases introduced in the previous similar methods. We demonstrate that a single-policy LaGarNet achieves flattening on four different types of garments in both real-world and simulation settings.