Rock-and-walk (video) is a method for dynamic, non-prehensile and underactuated object transport. Rock-and-walk is motivated by an interesting question in archaeology, how the giant rock statues of Easter Island (known as “moai”) were transported several hundred years ago, and a recent demonstration performed by archaeologists that it is possible to walk the statue by repetitive rocking. Here, we show that such manipulation capability can be acquired through reinforcement learning and deployed on real robot systems. We demonstrate successful object transport with the learned policy through a set of simulated and real-world experiments, performed with a robot arm and an aerial robot interacting with the object in a non-prehensile manner. While the object, which is in contact with a support surface, oscillates sideways passively under gravity, the robot uses the learned policy to move the object forward with a steady gait by regulating the mechanical energy and the posture of the object. Our experiments show that the learned policy can transport the object through unmodeled effects of terrain and perturbation.

Related Papers

1. A. Nazir, P. Xu, J. Rojas, and J. Seo, "Learning to Rock-and-Walk: Dynamic, Non-Prehensile, and Underactuated Object Locomotion through Reinforcement Learning," submitted to ICRA'22.
2. A. Nazir and J. Seo, "Passive Dynamic Object Locomotion by Rocking and Walking Manipulation," 2019 International Conference on Robotics and Automation (ICRA), Montreal, QC, Canada, 2019, pp. 7926-7932, doi: 10.1109/ICRA.2019.8794163.

• anaconda installed with Python 3.8
• gym to create environment for object transport
• trimesh to generate 3D object models
• pybullet simulator to resolve mechanics of interaction between the object and the support surface
• stable-baselines-3 to train the learning algorithm
• tensorboard (version 2.0.0) to monitor training progress

• ROS Melodic running on Ubuntu 18.04
• python-urx for control of UR10 Robot Arm
• rosserial to acquire data from Arduino 9 Axis Motion Shield

TODO

Clone the repository in your conda environment installed with the above dependencies.

git clone https://github.com/HKUST-RML/learn_rockwalk.git

Cone-Shaped Model in Simulation

cd learn_rockwalk/cone_simulation/Rock-Walk
pip install -e .

cd ..
python main_sim.py

Moai in Simulation

cd learn_rockwalk/moai_simulation/Rock-Walk
pip install -e .

cd ..
python main_sim.py

Type 'yes' when prompted to test the model.

For technical enquiry, please contact Abdullah Nazir (a.nazir[at]connect.ust.hk) and Pu Xu (pxuaf[at]connect.ust.hk)