Pose predicting KKT-loss for weakly supervised learning of robot-terrain interaction model

Code for RA-L publication

Requirements

Code was implemented for python 2.7.

Required libraries

numpy
scipy
Torch
tensorboardX

Dataset

Please download the dataset from https://drive.google.com/drive/folders/1qMwzeyThEgAincA_ldWhTd5rZHmGAJp7

Dataset is expected to be located in "../data/" folder.

Pretrained models

We provide all pretrained models in folder weights

Training

The rigid terrain prediction (experiment 1):

script train_d2rpz.py will train q_omega network
script train_s2d.py will train h_theta network
script train_s2d_rpz.py will train h_theta network by pose-predictiong loss backpropagation (Section 3.2)
script train_s2d_kkt.py will train h_theta network by kkt loss backpropagation(Section 3.1)

The flexible terrain prediction (experiment 2):

script train_sf2d.py will train h_theta network
script train_sf2d_rpz.py will train h_theta network by pose-predictiong loss backpropagation (Section 3.2)
script train_sf2d_kkt.py will train h_theta network by kkt loss backpropagation(Section 3.1)

Evaluation

Running the script eval.py will produce the same results as shown in the paper in Table 1

Running the script eval_flexible.py will produce the same results as shown in the paper in Table 2