jstmn / ikflow

Open source implementation to the paper "IKFlow: Generating Diverse Inverse Kinematics Solutions"

IKFlow

Normalizing flows for Inverse Kinematics. Open source implementation to the paper "IKFlow: Generating Diverse Inverse Kinematics Solutions"

Runtime curve for getting exact IK solutions for the Franka Panda (maximum positional/rotational error: 1mm, .572 deg):

Setup - inference only

git clone https://github.com/jstmn/ikflow.git && cd ikflow
poetry install --without dev
poetry shell

Setup - inference, training, and visualization

The following section outlines the setup procedures required to run the visualizer that this project uses. The only supported OS is Ubuntu. Visualization may work on Mac and Windows, I haven't tried it though. For Ubuntu, there are different system wide dependencies for Ubuntu > 21 and Ubuntu < 21. For example, qt5-default is not in the apt repository for Ubuntu 21.0+ so can't be installed. See https://askubuntu.com/questions/1335184/qt5-default-not-in-ubuntu-21-04.

Ubuntu >= 21.04

sudo apt-get install -y qtbase5-dev qtchooser qt5-qmake qtbase5-dev-tools libosmesa6 build-essential qtcreator
export PYOPENGL_PLATFORM=osmesa # this needs to be run every time you run a visualization script in a new terminal - annoying, I know

Ubuntu <= 20.x.y

(This includes 20.04 LTS, 18.04 LTS, ...)

sudo apt-get install -y qt5-default build-essential qtcreator

Lastly, install with pip:

git clone https://github.com/jstmn/ikflow.git && cd ikflow
poetry install --with dev
poetry shell

Getting started

> Example 1: Use IKFlow to generate approximate IK solutions for the Franka Panda

Evaluate a pretrained IKFlow model for the Franka Panda arm. Note that the value for model_name - in this case panda__full__lp191_5.25m should match an entry in model_descriptions.yaml

python scripts/evaluate.py --testset_size=500 --model_name=panda__full__lp191_5.25m

> Example 2: Use IKFlow to generate exact IK solutions for the Franka Panda

Additional examples are provided in examples/example.py. This file includes examples of collision checking and pose error calculation, among other utilities.

ik_solver, _ = get_ik_solver("panda__full__lp191_5.25m")
target_poses = torch.tensor(
    [
        [0.25, 0, 0.5, 1, 0, 0, 0],
        [0.35, 0, 0.5, 1, 0, 0, 0],
        [0.45, 0, 0.5, 1, 0, 0, 0],
    ],
    device=device,
)
solutions, _ = ik_solver.generate_exact_ik_solutions(target_poses)

> Example 3: Visualize the solutions returned by the fetch_arm__large__mh186_9.25m model

Run the following:

python scripts/visualize.py --model_name=fetch_arm__large__mh186_9.25m --demo_name=oscillate_target

Run an interactive notebook: jupyter notebook notebooks/robot_visualizations.ipynb

Notes

This project uses the w,x,y,z format for quaternions. That is all.

Training new models

The training code uses Pytorch Lightning to setup and perform the training and Weights and Biases ('wandb') to track training runs and experiments. WandB isn't required for training but it's what this project is designed around. Changing the code to use Tensorboard should be straightforward (so feel free to put in a pull request for this if you want it :)).

First, create a dataset for the robot:

python scripts/build_dataset.py --robot_name=panda --training_set_size=25000000 --only_non_self_colliding

Then start a training run:

# Login to wandb account - Only needs to be run once
wandb login

# Set wandb project name and entity
export WANDB_PROJECT=ikflow 
export WANDB_ENTITY=<your wandb entity name>

python scripts/train.py \
    --robot_name=panda \
    --nb_nodes=12 \
    --batch_size=128 \
    --learning_rate=0.0005

Common errors

1. GLUT font retrieval function when running a visualizer. Run export PYOPENGL_PLATFORM=osmesa and then try again. See https://bytemeta.vip/repo/MPI-IS/mesh/issues/66

Traceback (most recent call last):
  File "visualize.py", line 4, in <module>
    from ikflow.visualizations import _3dDemo
  File "/home/jstm/Projects/ikflow/utils/visualizations.py", line 10, in <module>
    from klampt import vis
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/klampt/vis/__init__.py", line 3, in <module>
    from .glprogram import *
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/klampt/vis/glprogram.py", line 11, in <module>
    from .glviewport import GLViewport
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/klampt/vis/glviewport.py", line 8, in <module>
    from . import gldraw
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/klampt/vis/gldraw.py", line 10, in <module>
    from OpenGL import GLUT
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/OpenGL/GLUT/__init__.py", line 5, in <module>
    from OpenGL.GLUT.fonts import *
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/OpenGL/GLUT/fonts.py", line 20, in <module>
    p = platform.getGLUTFontPointer( name )
  File "/home/jstm/Projects/ikflow/venv/lib/python3.8/site-packages/OpenGL/platform/baseplatform.py", line 350, in getGLUTFontPointer
    raise NotImplementedError( 
NotImplementedError: Platform does not define a GLUT font retrieval function

2. If you get this error: tkinter.TclError: no display name and no $DISPLAY environment variable, add the lines below to the top of ik_solvers.py (anywhere before import matplotlib.pyplot as plt should work).

import matplotlib
matplotlib.use("Agg")

Citation

@ARTICLE{9793576,
  author={Ames, Barrett and Morgan, Jeremy and Konidaris, George},
  journal={IEEE Robotics and Automation Letters}, 
  title={IKFlow: Generating Diverse Inverse Kinematics Solutions}, 
  year={2022},
  volume={7},
  number={3},
  pages={7177-7184},
  doi={10.1109/LRA.2022.3181374}
}

ikflow's Issues

I have a question about the negative logarithmic likelihood function during code training.

I have a question about the negative logarithmic likelihood function during code training. I found that the negative logarithmic likelihood function I was training had negative values.

Is this caused by my lack of normalization in processing the dataset? I did not modify the code for ikflow.

And when I looked at the code, I found that the
neg_log_likeli is defined by this code：

output, jac = self.nn_model.forward(x, c=conditional, jac=True)
zz = torch.sum(output**2, dim=1)
neg_log_likeli = 0.5 * zz - jac
loss = torch.mean(neg_log_likeli)

I didn't see the logarithmic operation for zz and jac here when watching Graph INN, nor did I see the softmax layer in the final output of the model. I have been troubled here for a long time, and I really hope to receive your explanation and assistance from the author.

new robot training

Thank you for your excellent work. If I want to train my own robot, what do I need to prepare

1. Add your own robot's URDF file
2. Modify the code to include it in robots. py
3. Modify the model_ Descriptions.yaml file
4. Generate Dataset
5. Training Model
6. Validate the model
   I have tried the above steps, but there are still many things that seem incorrect. Can you give me some guidance?