ContactKernels: Adaptive Proposals using Geometry-Conditioned Wrenches for Contact-Rich Manipulation

Motivation

Problem: Controller Uncertainty in Contact-rich Manipulation settings

Goal: Improve uncertainty using Contact information

Operational Space Control commanding ground truth hole pose, exhibits steady state error that prevents tight tolerance insertion

Text

Sampling-based Model Predictive Control

Sampling-based MPC provides necessary tooling for adaptation. MPPI/PS can be used to sample around

Sampling-based Model Predictive Control provides necessary language for adapting Operational Space Control, making it suitable for contact-implicit control for tasks such as insertion.

However, the default isotropic gaussian proposal kernel for sampling is inefficient

Behaviours that mimic the wiggly behaviour that humans employ before inserting a key into the hole  - while being informed by local contact

ContactKernels

1. Sampling-based MPC (MPPI/PS) for insertion task
2. Contact-conditioned proposal kernel

SMPC with Isotropic Gaussian Kernel

SMPC with proposed contact-conditioned Kernel

Contributions

System Overview

Results

ContactKernels enables

• Sample efficient adaptation to local contact
• Improved Insertion success across geometries
• Improved Insertion times

Results

ContactKernels enables

• Sample efficient adaptation to local contact
• Improved Insertion success across geometries
• Improved Insertion times

Results