thriftysnipe.mp4
This is the code from the winning submission to the 2021 Real Robot Challenge Phase 1.
See our final report from Phase 1 and videos of our policies in action.
The code is built off the rrc_example_package provided by the challenge organisers. For more details on how to use this code with Singularity and ROS 2, see the relevant documentation.
git clone https://github.com/RobertMcCarthy97/rrc_phase1.git
-
Install the pre-built Debain package.
-
Download our custom singularity image: user_image.sif. Otherwise, rebuild it yourself using 'user_image.def' and following these instructions.
-
Name the image
user_image.sif
.
To reproduce our results in simulation, train a control policy from scratch by running the following command:
singularity run /path/to/user_image.sif mpirun -np 8 python3 train.py --exp-dir='reproduce' --n-epochs=300 2>&1 | tee reproduce.log
-np
specifies the number of MPI processes that will be run in parallel. Expect inferior performance if less than 8 are used.
Details of all relevant arguments are found in rrc_example_package/her/arguments.py
.
Expect each epoch of training to take up to 10 mins.
First, download our winning 'pinching' model: final_pinch_policy.pt
To view our model performing the task in simulation:
-
Save the downloaded model as
rrc_example_package/her/saved_models/final_pinch_policy.pt
and execute the following command:singularity run /path/to/user_image.sif python3 demo.py
To deploy the model on the real robot:
-
Upload the model to the robot cluster following these instructions. The path to the model on the cluster should thus be
/userhome/final_pinch_policy.pt
. -
Ensure your configuration file links to this repository.
-
Login via ssh and call
submit
This should run the rrc_example_package/scripts/evaualte_stage1.py
script on the real robot.
This is a basic example for a package that can be submitted to the robots of the Real Robot Challenge 2021.
It is a normal ROS2 Python package that can be build with colcon. However, there are a few special files in the root directory that are needed for running/evaluating your submissions. See the sections on the different challenge phases below for more on this.
This example uses purely Python, however, any package type that can be built by colcon is okay. So you can, for example, turn it into a CMake package if you want to build C++ code. For more information on this, see the ROS2 documentation.
An example scripts using the simulation:
sim_move_up_and_down
: Directly uses theTriFingerPlatform
class to simply move the robot between two fixed positions. This is implemented inrrc_example_package/scripts/sim_move_up_and_down.py
.
To execute the examples, build the package and execute
ros2 run rrc_example_package <example_name>
For the challenge phases on the real robots, you need to provide the following files at the root directory of the package such that your jobs can executed on the robots:
run
: Script that is executed when submitting the package to the robot. This can, for example, be a Python script or a symlink to a script somewhere else inside the repository. In the given example, it is a shell script running a Python script viaros2 run
. This approach would also work for C++ executables. When executed, a JSON string encoding the goal is passed as argument (the exact structure of the goal depends on the current task).goal.json
: Optional. May contain a fixed goal (might be useful for testing/training). See the documentation of the challenge tasks for more details.
It is important that the run
script is executable. For this, you need to do
two things:
- Add a shebang line at the top of the file (e.g.
#!/usr/bin/python3
when using Python or#!/bin/bash
when using bash). - Mark the file as executable (e.g. with
chmod a+x run
).
When inside of run
you want to call another script using ros2 run
(as it is
done in this example), this other script needs to fulfil the same requirements.