Awesome

Pretrained DLCHOMP Networks for Manipulator Motion Planning

This repository provides pretrained networks for Deep-Learning-Based Covariant Hamiltonian Optimization for Motion Planning (DLCHOMP) of robotic manipulators for MATLAB®. These pretrained networks can output intermediate trajectory guesses for desired start to goal configurations in a given spherical obstacle environment.

Creator: MathWorks Development

Includes un-trained model: ❌

Includes transfer learning script: ✅
(See this page)

Supported Robots: ✅
(See this page)

Requirements

• MATLAB® R2024a or later
• Robotics System Toolbox™
• Deep Learning Toolbox™

Getting Started

Download or clone this repository to your machine and open it in MATLAB®.

Setup

Add path to the source directory.

addpath("src");

Download Pretrained Network

Use the code below to download the pretrained network for a supported robot. For a list of supported robots, see the table in the Metrics and Evaluation section. To demonstrate this workflow, download the pretrained DLCHOMP network for the KUKA LBR iiwa 7.

robotName = "kukaIiwa7";
data = helper.downloadPretrainedDLCHOMPForRobot(robotName);

Summarize Pretrained Network

Extract the pretrained network. You can summarize and analyze the network as needed to better understand the architecture of the network.

pretrainedNetwork = data.trainedNetwork;
summary(pretrainedNetwork);
analyzeNetwork(pretrainedNetwork);

If you only need the pretrained network and not the pretrained optimizer, skip to the DLCHOMP Details section. Otherwise, continue to obtain a pretrained optimizer.

Obtain Pretrained DLCHOMP Optimizer

Extract the pretrained optimizer.

pretrainedDLCHOMP = data.trainedDLCHOMP;

Predict Trajectory Using Pretrained DLCHOMP Optimizer

Use the pretrained optimizer to predict a trajectory between a start joint configuration and goal joint configuration in an obstacle environment.

% Extract obstacle environment that was used to train the DLCHOMP optimizer.
pretrainedDLCHOMP.SphericalObstacles = data.unseenObstacles;

% Predict a trajectory using the optimizer.
[optimWpts,optimTpts,solinfo] = optimize(pretrainedDLCHOMP,data.unseenStart,data.unseenGoal);

% Visualize results.
show(pretrainedDLCHOMP,optimWpts);

Create and Train DLCHOMP Optimizer for New Applications

To generate data and train a DLCHOMP optimizer to suit your application or task, follow the Train Deep-Learning-Based CHOMP Optimizer for Motion Planning example.

Train Custom DLCHOMP Optimizer Using Transfer Learning

Transfer learning enables you to adapt a pretrained DLCHOMP optimizer to your dataset. Follow these examples to create a custom DLCHOMP optimizer and train it for transfer learning to:

• A different number of waypoints in trajectory by following the Using Pretrained DLCHOMP Optimizer to Predict Higher Number of Waypoints example.
• A different spherical obstacle environment and/or a different set of CHOMP optimization options by following the Using Pretrained DLCHOMP Optimizer in Unseen Obstacle Environment example.

DLCHOMP Details

You can enhance the efficiency of optimization-based motion planning tasks by applying deep learning [1]. dlCHOMP is one such MATLAB® feature that utilizes a neural network initial guesser to provide an educated initial guess for a robot's intermediate start to goal trajectory, which is then optimized using the Covariant Hamiltonian Optimization for Motion Planning (CHOMP)[2] algorithm.

• Env: Spherical obstacle environment in which robot motion planning is to be performed. This is provided as a 4-by-N numeric input matrix to dlCHOMP. This input is the fed to the CHOMP and BPS Encoder modules.
• BPS Encoder: Basis point set encoder for the obstacle environment. The basis point set is a set of fixed points that are used to convert an arbitrary size obstacle environment into a fixed-size encoding vector. This encoding vector is then fed as an input to the Pretrained Network module along with the desired start and goal configurations of the robot.
• Pretrained Network: Feed-forward neural network that guesses an initial intermediate trajectory for a robot by taking a start configuration, an end configuration, and an obstacle environment encoding vector as inputs. An intermediate trajectory is a trajectory that does not include the start and goal configurations. More details on the architecture of this neural network can be seen in the Neural Network Details section below.
• CHOMP: Motion planning optimizer that uses the Covariant Hamiltonian Optimization for Motion Planning algorithm [2]. It takes the initial intermediate trajectory guess output of the Pretrained Network and the spherical obstacle environment Env as its inputs to then output an optimized start to goal trajectory.

Neural Network Details

This figure shows the architecture of the DLCHOMP neural network [2].

It takes a given motion task (a world obstacle encoding vector WB, a start configuration q1 and an end configuration qNt) to output an initial guess Q. Blocks of tapered Fully Connected Layers (gray) are combined like the DenseNet architecture [3] via skip-connections and concatenations (circular nodes).

Metrics and Evaluation

The test dataset for each pretrained network consists of 1000 data samples, identical to the validation dataset created during the network training phase. To augment the data, these 1000 test samples were flipped, capitalizing on the symmetric nature of the motion planning problem, resulting in a total of 2000 test data samples. The results in these tables are for these 2000 test data sample sets that were created for each robot.

Size and Accuracy Metrics

The table above defines the headers present in the table below:

CPU Time Metrics

The table above defines the headers present in the table below:

Note: Dual-arm YuMi® (abbYuMi) being the only two-armed robot in this list, takes much longer for optimization as compared to other robots due to the higher probability of self-collisions.

References

[1] J. Tenhumberg, D. Burschka and B. Bäuml, "Speeding Up Optimization-based Motion Planning through Deep Learning," 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan, 2022, pp. 7182-7189, doi: 10.1109/IROS47612.2022.9981717.

[2] N. Ratliff, M. Zucker, J. A. Bagnell and S. Srinivasa, "CHOMP: Gradient optimization techniques for efficient motion planning," 2009 IEEE International Conference on Robotics and Automation, 2009, pp. 489-494, doi: 10.1109/ROBOT.2009.5152817.

[3] S. J´egou et al., “The one hundred layers tiramisu: Fully convolutional densenets for semantic segmentation,” CoRR, vol. abs/1611.09326, 2016

Copyright 2024 The MathWorks, Inc.