Introduction
In the early 1970s, robotic manipulation was characterized by jerky, point-to-point movements. Most systems operated by moving one joint at a time or employing independent controllers for each motor, largely ignoring the complex physical interactions inherent in multi-link manipulators. Richard Paul’s seminal work introduced a dual-stage framework that revolutionized this approach, combining high-level planning with low-level dynamic execution to achieve smooth, coordinated motion.
Planning and Trajectory Generation
The planning phase determines the "how" of the movement. Utilizing visual input from a TV camera to locate objects, the system generates paths using third-degree polynomials. This mathematical approach ensures $C^2$ continuity, preventing sudden jumps in position, velocity, or acceleration, which results in "graceful" motion.
A critical component of the planner is its heuristic collision avoidance. To prevent the arm from "driving through the table," an "Up" heuristic ensures objects are lifted vertically before horizontal translation. When a potential collision is detected, the system dynamically modifies the trajectory by inserting intermediate points to reroute the arm safely.
Execution and Servoing
Once a trajectory is planned, the execution phase—or servoing—takes over. Moving beyond simple thermostat-like control, Paul implemented a system based on Lagrangian mechanics. This approach calculates the required motor torques by accounting for the entire arm's mass, the coupling between joints (where moving one link affects others), and the effects of gravity.
Due to the computational constraints of the PDP-6 computer, solving these complex dynamic equations in real-time was unfeasible. To circumvent this, the system pre-computes the necessary coefficients during the planning phase, allowing the servo program to run at a high frequency of 60 samples per second.
Kinematic Modeling
The system was developed for the "Scheinman Arm," a 6-DOF manipulator featuring both rotary and prismatic joints. Precise control required a rigorous mathematical description of the arm's geometry. The coordinate system established for each link allowed for the transformation of visual data into joint-space commands.
Results and Insights
The integration of dynamic compensation—accounting for gravity and centripetal forces—allowed the robot to avoid the "droop" and "overshoot" common in earlier systems. The results were significant for the era:
- Speed: The arm could traverse 40 inches in less than one second.
- Precision: Final positioning accuracy reached 0.02% (approximately 1/30th of an inch).
- Task Success: The system successfully performed multi-stage tasks, such as turning over a block, which required seamless integration of vision, path planning, and dynamic control.
Conclusion and Future Work
While groundbreaking, the system had limitations. The collision avoidance was primarily heuristic and suited for simple environments like tabletops. More complex scenarios involving overhanging obstacles would require advanced path-finding algorithms. Furthermore, the reliance on pre-computed dynamics made the system less adaptable to sudden environmental changes. Future work pointed toward "intercepting" moving objects, a task requiring real-time trajectory perturbation that pushed the limits of 1970s computing power.
Summary updated