Video

Please checkout the introduction video. It highlights some important keypoints in the paper!

Quick View

Cassie Blue autonomously traverses the Wave Field via the proposed reactive planning system. More Cassie experiments are still being conducted. More images will be updated!

What does This Reactive Planning System Contain?

This reactive planning system consists of a low-frequency (5 Hz) planning thread and a high-frequency (300 Hz) reactive thread (see here). The planning thread involves a multi-layer local map to compute traversability, a sub-goal finder, and an omnidirectional Control Lyapunov Function RRT*. Instead of a common waypoint-following or path-tracking strategy, the reactive thread copes with robot deviation while eliminating non-smooth motions via a vector field (defined by a closed-loop feedback policy) that provides real-time control commands to the robot's gait controller as a function of instantaneous robot pose.

What is Reactive and Why

A reactive planning replaces the concept of trajectory with a vector field arising as the gradient of a potential function. Therefore, instead of a common waypoint-following or path-tracking strategy, the reactive thread copes with robot deviation while eliminating non-smooth motions via a vector field (defined by a closed-loop feedback policy) that provides real-time control commands to the robot's gait controller as a function of instantaneous robot pose. In other words, the reactive planner automatically adjusts the control commands to handle any deviations the robot may incur!

Overall Pipeline for Autonomy

Illustration of how the various processes in the overall autonomy system are distributed and their computation frequencies. The larger boxes indicate various modules such as Data Acquisition, Planning, and Control. The smaller boxes are colored according to the processor that runs them.

Results

Both simulation and experimental results are provided. All the followings results are shown as GIFs, which are generated by extracting 300 frames from the original videos and then are played back by 10 FPS. The real videos with 1x speed can be found at here.

Experimental Results

More experiments are still being conducted! More visual will be added soon.

Cassie Blue on Undulating Terrain

The experimental terrain is the Wave Field is located on the North Campus of the University of Michigan. The Wave Field consists of sinusoidal humps with a depth of approximately 1.5 m from the bottom of the valleys to the crest of the humps; there is a second sinusoidal pattern running orthogonal to the main pattern, which adds 25 cm ripples peak-to-peak even in the valleys.

Wave Field on the North Campus of the University of Michigan

Simulation Results of an ALIP Model Robot on the Wave Field

The simulated robot is based on the ALIP model and accepts piece-wise constant inputs at the beginning of each step. The highlighted areas show the local maps being provided to the robot.

Simulation Results of an ALIP Model Robot in Synthetic, Noise, and Patchy Maps

The simulated robot is based on the ALIP model and accepts piece-wise constant inputs at the beginning of each step. The highlighted areas show the local maps being provided to the robot.

-- cluttered indoor scenes with obstacles and holes

-- noisy undulating outdoor terrains

-- high-level missions

Simulation Results of a Whole-body Cassie Simulator in a Perfect Map

To ensure the control commands from the reactive planning system are feasible for Cassie-series bipedal robots, we sent the commands via User Datagram Protocol (UDP) from ROS C++ to Matlab-Simmechanics, which simulates a 20 DoF of Cassie, using footfalls on the specified terrain. The simulator then sent back the pose of the simulated Cassie robot to the planning system to plan for the optimal path via UDP. The planner system successfully took the simulated Cassie to the goal without falling!

Required Libraries / Packages

1. Please install ROS Melodic.
2. Please install this package.
3. Please install planner_msgs.
4. Please install grid_map.

The structure should be catkin
└── src
         ├── CLF_reactive_planning_system (This package)
         ├── planner_msgs
         └── customized_grid_map

Datasets

All the datasets (bagfiles) are available on Google Drive: please download them here

Running

Using provided bagfiles:

1. Please download at least one bagfile from here.
2. catkin_make the package in the catkin folder.
3. source devel/setup.bash
4. roslaunch cassie_planning fake_robot.launch
5. roslaunch cassie_planning wavefield.launch
6. rosbag play -l -q bagfile.bag
7. To see the results, rosrun rviz rviz. You can directly open wavefield.rviz under the CLF_reactive_planning_system/rviz/ folder.

Onboard processing:

1. catkin_make the package in the catkin folder.
2. source devel/setup.bash
3. roslaunch cassie_planning wavefield.launch
   • ensure you have robot pose and map information published
4. To see the results, rosrun rviz rviz. You can directly open wavefield.rviz under the CLF_reactive_planning_system/rviz/ folder.

Note: More detailed explanation about how to use configuration files (*.yaml) with launch files will come soon.

Citations

The detail is described in: Efficient Anytime CLF Reactive Planning System for a Bipedal Robot on Undulating Terrain, Jiunn-Kai Huang and Jessy W. Grizzle. (PDF) (arXiv)

@article{huang2021efficient,
      title={{Efficient Anytime CLF Reactive Planning System for a Bipedal Robot on Undulating Terrain},
      author={Jiunn-Kai Huang and Jessy W. Grizzle}},
      year={2021},
      journal={arXiv preprint arXiv:2108.06699},
      primaryClass={cs.RO}
}