In this repo, we implement a general-purpose exploration + task and motion planning agent from perceptual input (RGBD Images) and natural language goals. This repo builds off the following two papers:
Long-Horizon Manipulation of Unknown Objects via Task and Motion Planning with Estimated Affordances
Visibility-Aware Navigation Among Movable Obstacles
Please contact [email protected] before attempting to use it for your own research.
Clone the repo and its submodules (may take a while due to the large amount of logged data):
git clone [email protected]:Learning-and-Intelligent-Systems/open-world-tamp.git
cd open-world-tamp
git submodule update --init --recursive
Install the python dependencies. If possible, install using python3.8 as that appears to be the only python version that supports all of the perceptual dependencies:
$ python -m pip install -r requirements.txt
If you get errors when installing detectron, you may need to modify your paths. Make sure to switch out 11.4
for your current cuda version.
export CPATH=/usr/local/cuda-11.4/targets/x86_64-linux/include:$CPATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.4/targets/x86_64-linux/lib:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda-11.4/bin:$PATH
Build FastDownward:
./tamp/downward/build.py
Compile IKFast:
cd pybullet_planning/pybullet_tools/ikfast/<robot-name>
python setup.py
If you're looking to use the segmentation network with the --segmentation
flag, you will need to download the pretrained UCN checkpoint from
here and place the checkpoints folder in vision_utils/ucn/data
Command line arguments are used to specify the robot, goal, and simulated world.
python run_planner.py --robot=pr2 --goal=all_green --world=problem0 -v
--simulated
Specifies if the run is simulated or in the real-world. True by default.
--world
If simulated, this argument tells OWT how to set up the simulated world
--segmentation
Specified if the run uses segmentation networks instead of ground truth segmentation from the simulator. False by default.
--robot
Specifies the robot you are using. Current options are pr2
, panda
, movo
--goal
Specifies the objective of the robot. For example, all_green
instructs the robot to place all movable objects on a green region.
-v
Is a flag to visualize the resulting plan. False by default.
--real
Is a flag that specifies if mobile-base exploration should proceed manipulation planning
--base_planner
Is a flag that specifies the algorithm to use for mobile-base exploration
This is only a subset of the available segmentation flags. See code for more.
Combine mobile-base exploration with fixed-based manipulation by calling the planner with the following flags
--exploration
Toggles exploration
--base_planner
Selects the planner to use for exploration. Default is VA*, but more advanced planners can also be used.
Run the automated tests with the following command
pytest tests/
Run a coverage test with the following command. You can see the coverage report by opening htmlcov/index.html in a browser.
pytest --cov-config=.coveragerc --cov=. --cov-report html tests/