Basic adaptive control theory and its application to robot control-II.

Masayoshi TomizukaJournal of the Robotics Society of Japan 1985 · DOI

Abstract

This paper explores the application of discrete-time adaptive control theory, specifically the Recursive Least Squares (RLS) method (小2乗法), to robotic manipulators. It addresses the challenges of controlling non-linear, time-varying systems by implementing real-time parameter estimation to handle varying payloads and configurations.

Introduction: The Challenge of Robotic Control

Robotic manipulators are notoriously difficult to control with traditional fixed-gain controllers because their dynamics are highly non-linear and time-varying. As a robot moves, its effective inertia changes based on its posture, and picking up a heavy object (payload) fundamentally alters its physical parameters.

In 1985, the shift toward digital computer control necessitated "discrete-time" versions of these algorithms. The challenge was to create a system that could "learn" these changing parameters on-the-fly and update the control law fast enough for smooth motion, all while ensuring the system remained stable.

Methodology: RLS and Layered Control

The solution involves two main components: a Parameter Adaptation Algorithm (PAA) and a Pole-Placement Controller.

Parameter Adaptation via Recursive Least Squares (RLS)

The paper focuses on Recursive Least Squares (RLS). Unlike batch processing, RLS updates its estimate of the robot's parameters (like mass and friction) with every new sensor sample. A key feature is the "forgetting factor," which allows the algorithm to prioritize recent data over old data—essential for tracking parameters that change as the robot moves.

Hierarchical Control Structure

The method is applied to multi-axis robots using a hierarchical approach, where a fast adaptive velocity loop is nested inside a position control loop. This simplifies the overall design by allowing the inner loop to handle joint dynamics while the outer loop manages trajectory tracking.

Block diagram of the nested position and adaptive velocity control system.

Once parameters are estimated, the controller solves the Bezout (or Diophantine) identity to specify exactly where the "poles" of the closed-loop system should be, defining the desired speed and damping. This approach cancels out inter-joint interference (Coriolis and centrifugal forces) and gravity effects, making each joint behave like an independent, linear system.

Key Results: Achieving Payload Invariance

The performance was evaluated through numerical simulations of a robotic manipulator under three distinct payload conditions: 0 kg, 0.5 kg, and 1.36 kg. Simulations showed that while a controller without adaptation performed poorly when the payload deviated from the design point, the adaptive controller maintained high performance.

With the adaptive controller engaged, the robot successfully tracked desired trajectories across all payload levels with minimal error. The adaptive law successfully compensated for gravity and joint coupling, allowing the robot to maintain precise path following regardless of its posture.

Simulation results showing consistent tracking performance across different payload masses (0kg, 0.5kg, 1.36kg) with adaptive control.

Discussion and Limitations

Despite the theoretical success, several practical challenges were identified:

  • Computational Burden: In 1985, the matrix inversions required significant CPU power, suggesting a need for parallel processing or custom VLSI hardware.
  • The "PE" Condition: Accurate estimation requires "persistently exciting" (PE) input signals; simple or static movements can lead to estimation drift.
  • Robustness: Real-world sensor noise and unmodeled high-frequency dynamics (like joint flexibility) remain challenges that require further robustification of the adaptive laws.

Conclusion

This foundational work demonstrates that discrete-time adaptive control, particularly RLS-based pole-placement, provides a robust framework for managing the complex dynamics of robotic manipulators. By enabling real-time adaptation to physical changes, these methods ensure consistent performance across a wide range of operating conditions.

Summary updated

Abstract

小2乗 法(recursive least squares, RLS)
BibTeX
@article{tomizuka1985basic,
  author = {Masayoshi Tomizuka},
  title = {{Basic adaptive control theory and its application to robot control-II.}},
  journal = {Journal of the Robotics Society of Japan},
  year = {1985},
  doi = {10.7210/jrsj.3.610},
}
Figure