Boston Dynamics Spot Model for Ignition Gazebo

[peci1]

Let's cooperate in creating Boston Dynamics Spot simulation model! Leave here a message if you're interested in cooperating with us (CTU-CRAS-NORLAB), or send pull requests against brach spot in this fork if you already have code you'd like to share.

[bfotheri]

Little late here but I'm from team MARBLE with CU Boulder and we are going to be using the spot model for the finals. Clearpath robotics has an awesome URDF of spot found here: https://github.com/clearpathrobotics/spot_ros/blob/master/spot_description/urdf/spot.urdf.xacro

I will start with the URDF and work to port it to an SDF, and add the necessary ROS plugins and ignition files. I'm starting in earnest here in a week or two. I can post updates here periodically and will look for opportunities to collaborate during this process.

[peci1]

Good timing, I wanted to start working on the Spot model this or the next week :) So let's see who's faster - just be sure to write an update here as soon as you have something so that we don't duplicate our work.

Be sure to check out https://taras-borovets.medium.com/simulation-tools-for-boston-dynamics-spot-28a10d88e325 .

Do you already have the Spot hardware? Because we've just found out that due to administrative obstacles, our Spots will probably not arrive before the virtual model submission deadline.

According to the model submission wiki entry, the real-robot validation data need to be provided only for the first sensor config of each robot. For other configs, the wiki doesn't state they would need the bagfiles and recording of how the sensors work on the real robot. So our idea is that you (or Costar, who we asked for cooperation) would record the first sensor config validation data, and we would submit our (planned) sensor config "on top" of the first sensor config (which should not need putting the sensor rack on the real robot and recording data, but I'll verify with DARPA).

[peci1]

So, the first version of the Spot model is there! d02c4f9

spot.mp4

It is on branch spot in this repo. If anyone wants write access to ease the cooperation, let me know.

So far, I've got URDF, SDF with inertias, collisions, visuals and motors (effort joint controllers) and the respective ROS-Ign bridges to allow control from ROS. The motors can be controlled individually or in batches (via a custom plugin). It is largely based on https://github.com/chvmp/spot_ros/tree/gazebo to ease future integration of CHAMP.

I named the "bare" robot bosdyn_spot as I considered it weird to assign it the name of one of the teams when it will be done in cooperation. I also started an example ctu_cras_norlab_spot_sensor_config_1 robot to show how a sensor config can utilize the "bare" robot package to minimize code duplication. I really like what can be achieved with the power of xacro/ruby/roslaunch/bash, but I understand it might seem complicated at first sight. Don't hesitate to ask about it or suggest improvements or simplifications :)

It might also seem unusual that this packages inverts the direction of "authority" regarding SDF and URDF files. In most submitted models, there is a hand-crafted SDF, which is then somehow transformed to URDF. I chose the other way round since it seems more powerful. So the "authoritative" source of all 3D models is the Xacro file, which is then converted to URDF, which is then converted to SDF. The SDF requires some polishing after going through this chain, but that's an easy sed job. After you edit the Xacro files, just call scripts/update_robot_sdf from the source space and the SDF file will get regenerated.

Now the tough part comes, integrating the gait controller via CHAMP, maybe stairs climber via TOWR :)

Also, what I'm missing and will be unable to do without the real robot is getting some kind of specs of the cameras. Does anyone have information about what kind of cameras is it? I've read somewhere that it is grayscale stereo, is that correct? Could you maybe share a short bagfile of all topics the robot publishes via spot_ros? That would also help getting an idea on how to make the simulation closer to real behavior.

[grassjelly]

hey @peci1 this might be a good start on your stairs climbing integration https://github.com/hgaren/aliengo_delivery . They got the controller to handle stairs using the Aliengo config.

[peci1]

Need to tune the PIDs a bit :)

spot.mp4

[peci1]

Better!

spot2.mp4.mp4

[peci1]

To get state estimation working, we'll need PR osrf#834 to be accepted to be able to use contact sensors.

[peci1]

Last commit on the spot branch ( https://github.com/ctu-vras/subt/tree/spot ) contains a lot of improvements making the control of the robot finally practical using CHAMP, and basic state estimation and odometry work too! I'll write up a summary of how to control the robot and how to build a custom sensor config on top of it later today.

@bfotheri would you be able to provide some sample Bag file until the model submission deadline, or even better record all the model validation tasks? Our Spots are still blocked by the administration (but we're working on it).

[peci1]

The base Spot model (bosdyn_spot) has been finished! It walks on flat surface with speeds over 1 m/s when running straight! Terrain has not yet been tested, but as it uses an open-loop controller, it can't work well. I think the feedback loop and improved controller are left as a practice to the reader :-D

I've also equipped the robot with an IMU, 5 mono cameras and 5 depth cameras. According to the docs I've found, this should be it. I'm still missing any (even short) bag files from the real robot that would help me get the sensory model a bit more realistic.

Basic odometry is also working (but still waiting for osrf#834 to be supported on the official worlds). To test it before this update gets merged, launch LANG=C ign launch -v4 src/subt/submitted_models/bosdyn_spot/launch/example.ign robotName:=X1 ros:=true champ:=true.

The only thing that's missing now is finishing up the specifications.md file to be up-to-date, but It's already pretty late here, so I'll leave that to Sunday. When that's finished, I'll open a PR to merge this robot in SubT virtual repo.

This means that other teams (@bfotheri and CoSTAR AFAIK) can start working on adding their own sensor configs. An example is shown in ctu_cras_norlab_spot_sensor_config_1 (also on the spot branch of this repo). You can mostly copy-paste it and rename it to your desired name. The CTU-CRAS-Norlab sensor config contains extra goodies to re-paint the spot to our blue color. If you do not need to re-texture the robot, you can leave out the meshes folder and a few lines in update_robot_sdf and print_robot_urdf.

The SDF model development process is "inverted" compared to the process used by majority of SubT models. SDF is not the king here. The king is Xacro. From the set of Xacro files, both RViz model URDF and Gazebo SDF are generated. The URDF is generated completely at runtime so you won't find it anywhere in the repo. But it's enough to call print_robot_urdf name:=X1 and you'll get the URDF. With SDF, the model.sdf file is needed to be stored on disk. So the workflow is such that you edit the Xacro files and then call update_robot_sdf. This script updates the model.sdf file. All manual changes to the SDF file are lost each time, so it's really better to make all changes in the corresponding Xacro sections. You also have to call this script if nothing changed in you sensor config, but the base package bosdyn_spot had some Xacro update. This is not done automatically.

In addition to the Xacro files, there are YAML configs in config/model. These form a hierarchical set of configuration files that affect the Xacro files (e.g. friction coefficients can be set in the config). The config of bosdyn_spot is always loaded first, and then follow the configs from your sensor config which can add or override any value. There is common.yaml file which contains all common parameters. Then there is urdf.yaml which contains a set of parameters that are used only when generating the RViz URDF. Finally, there's ign.yaml, which is used only when updating the model.sdf file.

To get the basic walking controller and odometry, clone https://github.com/ctu-vras/subt_spot to your workspace and install its dependencies. Once you do that and launch subt_spot/launch/champ.launch name:=X1, the robot will lower on its legs to a walking-suitable position and will await commands on cmd_vel and body_pose. For easy testing with a gamepad or keyboard, try https://github.com/chvmp/champ_teleop .

spot.mp4

[peci1]

Pull request for the base robot and our 2 sensor configs has been already been opened! osrf#841

[peci1]

https://app.ignitionrobotics.org/OpenRobotics/fuel/models/bosdyn_spot

[peci1]

We've finally got our hands on a real Spot so we recorded the validation data: osrf#913 .