Tutorial 5: Playing with Gravity
In this tutorial we will explore how gravity affects robot motion, how it can be compensated for in the controller, and what interesting behaviors emerge as a result.
Background
As introduced in the previous tutorial, each joint can be modeled as a virtual spring-damper system. With this analogy in mind, consider what happens when you hang a mass from a spring: gravity pulls the mass downward, displacing it from the neutral position until the spring force reaches equilibrium with gravity.
At rest, this balance is described by:
m · g = k · Δx
The same effect occurs on a robot arm. When a joint holds a link against gravity, the gravity torque acts like an external disturbance, causing a steady-state position error that the PD controller alone cannot fully eliminate. The following demos will let you observe and compare this behavior directly on the hardware.
Experiment 1 - Without Gravity Compensation
Open a terminal and navigate to the tutorial repository:
cd ulixarm_ros2
You can open the project with your favourite code editor. To open it in VSCode run
code .
Open the file at the following path:
~/ulixarm_ros2/src/ulixarm_description/src/ulixarm_moveit_config/config/ulixarm.ros2_control.xacro
In the <hardware> tag, set use_gravity_compensation to false and ensure use_free_floating is also set to false:
<hardware>
<param name="use_gravity_compensation">false</param>
<param name="use_free_floating">false</param>
</hardware>
Build the packages:
colcon build
Launch the demo:
ros2 launch demos moveit_rviz.launch.py
In RViz, drag the interactive marker to a new pose and click Plan and Execute. For a more detailed tutorial on how to interact with the robot from RViz, consider following the tutorial: Moving the Robot (MoveIt 2 + RViz).
The orange robot represents the goal pose. The textured robot represents the real pose of the robot, read from the joint encoders.
Observe the difference between the two and reflect on the following questions:
How large is the position error between the goal and the real pose?
Based on what you learned in Tutorial 4, how does increasing
Kponjoint2affect the error? Does the error ever fully vanish?
Experiment 2 - With Gravity Compensation
As introduced in Tutorial 4, the MIT controller accepts a feed-forward torque input tau_ff. This term allows the motor to apply a torque independently of the position and velocity error, making it an ideal channel for gravity compensation.
To compensate for gravity, tau_ff is set to the gravity torque vector g(q), computed at each control loop via pinocchio::computeGeneralizedGravity(q).
Note that g(q) depends on the current robot configuration. Consider the difference in effort required at joint2 when the arm is stretched vertically versus horizontally, the gravity torque changes significantly with pose.
In the <hardware> tag, set use_gravity_compensation to true and ensure use_free_floating remains false:
<hardware>
<param name="use_gravity_compensation">true</param>
<param name="use_free_floating">false</param>
</hardware>
To apply the changes, build the packages:
colcon build
Launch the demo:
ros2 launch demos moveit_rviz.launch.py
Plan and execute a trajectory as before, then reflect:
Has the position error improved compared to Demo 1?
What factors might still limit the accuracy of gravity compensation?
No model is perfect. The inertial parameters (masses, center of mass positions) are estimated from the CAD model and carry some error. Additional unmodeled effects, such as joint friction, which varies with temperature and wear — further limit how accurately gravity can be compensated.
Experiment 3 - Free Floating
What happens if we stop commanding the robot to hold a fixed position and instead only compensate for gravity? The joint torques balance gravity exactly, leaving no net force acting on the robot, it behaves as if it were floating in zero gravity.
In the <hardware> tag, set both flags as follows:
<hardware>
<param name="use_gravity_compensation">true</param>
<param name="use_free_floating">true</param>
</hardware>
To apply the changes, build the packages:
colcon build
Launch the demo:
ros2 launch demos moveit_rviz.launch.py
You can now move the robot freely with your hand. When you release it, the robot will hold whatever pose it was left in, with no position command driving it. Enjoy the experiment!