The Demo
Project Overview
This project implements a Reactive Controller for the 7-DOF Franka Emika Panda robot. This system uses Artificial Potential Fields (APF) to calculate control commands made from scratch in real-time (100Hz). This approach allows the robot to instantly adapt to dynamic environments, dodging obstacles while reaching for the target position.
Background
- The Goal: The target position naturally pulls the robot’s end effector.
- The Obstacles: Obstacles generate a repulsive “energy hill” that pushes the robot away. This force activates only when the robot breaches a safety threshold. It grows exponentially as the robot gets closer to an obstacle, creating a virtual bumper.
- The Result: The robot acts as a particle moving through this energy landscape, naturally flowing around obstacles to reach the goal without explicit path planning.
Challenge & Resolution
The Challenge: Instability and Oscillation A major difficulty in implementing APF is tuning the gain parameters (zeta and eta).
- High gains allowed the robot to overpower the repulsion field and reach the target, but caused violent high-frequency oscillation when forces conflicted.
- Low gains resulted in smooth motion but left the robot “stuck” in local minima, unable to push through the safety threshold of the obstacle or overcome its own posture constraints.
The Resolution: Global Damping and Gain Scheduling To stabilize the system without sacrificing strength:
- Global Damping: Implemented a velocity scaling factor. This allowed for high stiffness gains while dissipating the kinetic energy that caused vibrations.
- Safety Clamping: Capped the maximum repulsive force to prevent mathematical singularities when the robot operated extremely close to obstacle surfaces.
- Dynamic Braking: Designed a non-linear braking profile that smoothly decelerated the robot within the final 2cm of the target.
Simulation Results
The simulation, built in MuJoCo and wrapped in ROS 2, demonstrates the robot successfully navigating a “glancing blow” scenario. The robot curves around a central obstacle to reach a target, adjusting its path instantaneously as the obstacle pushes against the repulsion field.
Key Learning Outcomes
- Artificial Potential Fields (APF): Developed a strong intuition for designing reactive control. Learned to shape robot behavior mathematically by constructing virtual energy landscapes—defining attraction basins for goals and exponential repulsion fields for safety.
- Tuning Non-Linear Systems: Explored the trade-off between Stiffness (Accuracy) and Stability (Smoothness). Discovered that high-gain systems require robust “shock absorbers” (global damping) to prevent high-frequency oscillations when conflicting forces meet.
Citation
https://www.cs.cmu.edu/~motionplanning/lecture/Chap4-Potential-Field_howie.pdf