plainmp

The project is licensed under the BSD 3 License (see LICENSE), except for the code in cpp directory which is licensed under the MPL2.0 (see cpp/LICENSE-MPL2).

plainmp provides:

• Fast sampling-based motion planning (e.g., less than 1ms for moderate problems using RRTConnect)
• Collision-aware inverse kinematics (IK) solver
• Motion planning/IK for various models (e.g. movable base, dual-arm, object attachment)
• Flexible framework for defining various robot model and motion planning problems
• Collision checking for primitives (sphere/box/cylinder...) and/or point cloud vs. robot
• (Beta) Sampling-based constrained motion planning solver (e.g., whole-body humanoid)
• (Beta) SQP-based constrained motion planning (will be used as smoother for sampling-based planner)

Note that plainmp currently heavily relies on approximations of robot body by spheres.

The TODO list is

• Speed up IK by reimplementing current Python/C++ mix into pure C++
• Auto-generate collision spheres from URDF instead of manual sphere definitions

Related/depeding projects:

• plainmp is a rewrite of my previous projects scikit-motionplan and tinyfk to achieve 100x speedup
• plainmp depends on OMPL (with unmerged PR of ERTConnect) for SBMP algorithms
• plainmp deponds on scikit-robt framework for visualization and planning problem specifications
• benchmark with VAMP (the world fastest motion planner as of 2024 to my knowledge) shows that
  • AMD Ryzen 7 7840HS (256-bit AVX) VAMP is faster (1.3x, 1.1x, 4.8x, 12.5x)
  • ARM Neoverse-N1 (128-bit NEON) both seems to be comparable (0.53x, 0.41x, 2.2x, 4.8x)
  • x-s are time ratio plainmp/VAMP for 4 different settings with resolution of 1/32

Performance example

panda dual bars: median 0.17 ms | panda ceiled dual bars: median 0.65 ms | fetch table: median 0.62 ms

* resolution is isotropically set to 0.05 rad or 0.05 m with box motion validator

* On my laptop (AMD Ryzen 7 7840HS)

The plots are generated by the following commands:

python3 example/bench/panda_plan.py  # panda dual bars
python3 example/bench/panda_plan.py --difficult  # panda ceiled dual bars
python3 example/bench/fetch_plan.py  # fetch table

installation and usage

sudo apt-get install libspatialindex-dev freeglut3-dev libsuitesparse-dev libblas-dev liblapack-dev  # for scikit-robot
sudo apt install libeigen3-dev libboost-all-dev libompl-dev # plainmp dependencies
pip install scikit-build
git submodule update --init
pip install -e . -v

Then try examples in example directory with --visualize option. Note that you may need to install the following for visualization:

pip uninstall -y pyrender && pip install git+https://github.com/mmatl/pyrender.git --no-cache-dir

Troubleshooting

• Segmentation faults and other C++ runtime errors may occur when multiple OMPL versions are present - typically when installed via both from Ubuntu and ROS. To temporarily resolve this, disable ROS workspace sourcing in your shell or remove either OMPL installation.

How to add a new robot model

• (step 1) Prepare a URDF file. Note that robot_description package might be useful.
• (step 2) Implement a new class inheriting RobotSpec class in python/plainmp/robot_spec.py.
• (step 3) Write yaml file defining urdf location/collision information/control joints/end effector in (see example yaml files).
• NOTE: In step 3, you need to manually define the collision spheres for the robot (This is actually tedious and takes an hour or so). For this purpose, a visualizer script like this might be helpful to check the collision spheres defined in the yaml file. The output of the this visualizer looks like figure below.

Note on motion validator of motion planning

We provides two types of motion validator type box and euclidean.

• The box type split the motion segment into waypoints by clipping the segment with the collision box. You need to specify box widths for each joint.
  • e.g. bow_width = np.ones(7) / 20 for panda robot
• The euclidean type is instead split the segment by euclidean distance of resolution parameter.
  • e.g. resolution = 1/20 for panda robot
• see problem.py for how to specify the motion validator.