generative-graphIK

Code for paper on Generative Graphical Inverse Kinematics.

Installation

Install matching versions of PyTorch and PyTorch-Geometric. We used torch-1.10.1 and torch-geometric-2.0.4 but other versions should work as well.

After installing the above:

pip install -e .

Generate a dataset and train

./train.sh <yourmodelname> <yourdatasetname>

See ./generative-graphik/generative_graphik/args/parser.py for more details on data generation and model parameters.

Modifying data generation

To modify the training data, modify lines 29-35 of train.sh.

To train on specific robots:

    python -u ${SRC_PATH}/generative_graphik/utils/dataset_generation.py \
        --id "${DATASET_NAME}" \
        --robots ur10 kuka panda lwa4d lwa4p \
        --num_examples 512000 \
        --max_examples_per_file 512000 \
        --goal_type pose \
        --randomize False

To train on random robots of DOFs 6 and 7:

    python -u ${SRC_PATH}/generative_graphik/utils/dataset_generation.py \
        --id "${DATASET_NAME}" \
        --robots revolute_chain \
        --dof 7 6 \
        --num_examples 512000 \
        --max_examples_per_file 512000 \
        --goal_type pose \
        --randomize True \
        --randomize_percentage 0.4

Modifying hyperparameters

To modify the model parameters, modify lines 43-77 of train.sh.