Implicit Neural-Representation Learning for Elastic Deformable-Object Manipulations

TL;DR

Deformable object manipulation is challenged by infinite degrees of freedom and partial observations. The authors propose INR-DOM, an implicit neural representation framework that learns occlusion-robust state embeddings by reconstructing complete surfaces as signed distance functions, coupled with a two-stage training pipeline. The first stage pre-trains a partial-to-complete representation; the second stage fine-tunes with reinforcement learning and contrastive learning to produce task-relevant representations and exploitable policies. In simulation and real-world experiments with a Franka Panda, INR-DOM achieves superior reconstruction accuracy and higher task success rates, demonstrating effective sim-to-real transfer for elastic DOM tasks.

Abstract

We aim to solve the problem of manipulating deformable objects, particularly elastic bands, in real-world scenarios. However, deformable object manipulation (DOM) requires a policy that works on a large state space due to the unlimited degree of freedom (DoF) of deformable objects. Further, their dense but partial observations (e.g., images or point clouds) may increase the sampling complexity and uncertainty in policy learning. To figure it out, we propose a novel implicit neural-representation (INR) learning for elastic DOMs, called INR-DOM. Our method learns consistent state representations associated with partially observable elastic objects reconstructing a complete and implicit surface represented as a signed distance function. Furthermore, we perform exploratory representation fine-tuning through reinforcement learning (RL) that enables RL algorithms to effectively learn exploitable representations while efficiently obtaining a DOM policy. We perform quantitative and qualitative analyses building three simulated environments and real-world manipulation studies with a Franka Emika Panda arm. Videos are available at http://inr-dom.github.io.

Figures (10)

• Figure 1: A capture of deformable object manipulation task that requires disentangling elastic bands between two poles. The deformable and stretchable nature of bands increases the complexity of state representation. Further, the $360^\circ$ twists create self-occlusions, significantly reducing the consistency of state embeddings. Our INR-DOM effectively captures the occlusion-robust implicit representation of the bands and efficiently generates real-world applicable manipulation policies.
• Figure 2: An overview of INR-DOM framework that aims to train the occlusion-robust state representation encoder $\Phi_\phi$, parameterized by $\phi$, of deformable objects (DOs) as well as the manipulation policy $\pi$. The training framework consists of two stages: 1) The first stage pre-trains a PointNet-based partial-to-complete variational autoencoder $(\Phi_\phi, \Psi)$ that embeds a partial point cloud ${\bf p}$ of a target DO into a latent embedding ${\bf z}$ and recovers the parameters $\boldsymbol{\theta}$ of an implicit signed distance field (SDF) network $\Omega_\theta$. This stage predicts full geometries leveraging two loss functions $\mathcal{L}_{\text{SDF}}$, $\mathcal{L}_{\text{skel}}$, along with three regularization loss functions: $\mathcal{L}_{\text{KL}}$, $\mathcal{L}_{\text{weight}}$, and $\mathcal{L}_{\text{cns}}$. 2) The second stage then improves the task-relevant representation power of the encoder $\Phi_\phi$ by jointly optimizing reinforcement learning (blue) with the loss $\mathcal{L}_{RL}$ and the contrastive learning (red) with the loss $\mathcal{L}_{\text{infoNCE}}$.
• Figure 3: Examples of randomly twisted and stretched rubber bands in simulation. The red points represent partial point clouds.
• Figure 4: Examples of three deformable-object manipulation environments: sealing, installation, and disentanglement.
• Figure 5: Comparison of point-cloud reconstruction performance for both seen and unseen types of partially observable rubber bands. The blue and red bars represent the reconstruction errors measured by chamfer distance (CD) and earth mover's distance (EMD), respectively. Note that INR-DOM$^{-skel}$ refers to a variant of INR-DOM that was not pre-trained using the $\mathcal{L}_\text{skel}$ loss.