TL;DR
This paper presents a novel adaptive compliant control strategy for robot arms using Artificial Neural Networks (ANNs). The approach effectively decouples the robot's trajectory control from its compliant behavior, allowing it to handle both precise movements and unexpected disturbances like collisions with flexibility. This offers a simpler, hardware-agnostic solution for safer and more robust robot arm operation.
What Problem Does This Paper Address?
Controlling robot arms, especially in scenarios involving human interaction or unpredictable environments, presents a significant challenge. While linear control strategies are often sufficient, they struggle to compensate for inherent nonlinearities like friction and backlash. Existing nonlinear control techniques typically require precise knowledge of the robot's dynamics, which is often difficult to obtain. Furthermore, ensuring safety during human-robot interaction necessitates compliant behavior, allowing the robot to absorb impacts and avoid injury. This paper addresses the need for a robust, adaptive, and compliant control system that can handle nonlinear dynamics and external disturbances without requiring detailed prior knowledge of the system or specialized hardware.
Methods
The proposed control strategy is based on a computed torque law, which is a standard approach for robot control. However, this paper introduces a novel ANN-based feedforward controller to approximate the complex nonlinear dynamics of the robot arm. This ANN is trained to learn the system's dynamics, allowing for adaptive control. This feedforward component is combined with a proportional-derivative (PD) feedback loop.
The key innovation lies in the decoupling of these two components:
- ANN Feedforward Controller: Approximates the robot arm's nonlinear dynamics and can adapt to long-term changes in the system (e.g., wear and tear, changes in payload). This is achieved using a multilayer ANN.
- PD Feedback Loop: Provides short-term disturbance rejection and compliant behavior. This loop is tuned to absorb instantaneous impacts, such as collisions, without destabilizing the system.
The ANN architecture includes both smooth and nonsmooth activation functions to better handle nonlinear friction effects. The learning algorithm uses a standard error backpropagation method with an added "forgetting term" to ensure bounded weights and stable behavior.
Datasets
Not discussed in the paper. The evaluation was performed in a physical simulation environment (LPZRobots).
Key Insights and Ideas
- Decoupling for Simplicity and Effectiveness: The ability to independently tune the ANN for adaptive dynamics compensation and the PD loop for compliance is a significant advantage. This simplifies the design and tuning process.
- ANNs as Universal Approximators for Robot Dynamics: The paper demonstrates the practical utility of ANNs in approximating complex, unknown robot arm dynamics, a key step towards model-free control.
- Handling Both Long-Term Adaptation and Short-Term Compliance: The combined approach effectively addresses both gradual changes in the robot's dynamics (learned by the ANN) and sudden external disturbances (handled by the PD loop).
- Hardware Agnosticism: The proposed control scheme does not require additional hardware like force sensors, making it more broadly applicable.
Key Results
The paper presents simulation results comparing the ANN-based compliant controller with a conventional PID controller.
- Sinusoidal Reference Tracking: The ANN controller achieved comparable or better tracking performance than the PID controller for sinusoidal reference signals, as shown in Figure 3.
- Disturbance Rejection:
- Instantaneous Push Force: The ANN controller, with its PD loop, effectively compensated for an instantaneous push force at t=2.5s, demonstrating compliant behavior. The learning curve showed minimal weight changes, indicating the PD loop handled the disturbance.
- Change in Payload: When the weight of a carried mass changed at t=5.0s, the ANN controller successfully adapted to the new dynamics, eliminating tracking errors. The PID controller, with low feedback gains, struggled to recover from this change.
- Learning Convergence: The ANN successfully learned the robot's dynamics within approximately 0.5 seconds, as indicated by the learning curves in Figure 4.
Limitations and Future Work
- Simulation-Based Evaluation: The current results are based on physical simulation. The authors acknowledge that implementing the controller on a real robot arm is necessary for further validation.
- Potential for Oscillatory Response: The paper notes that while higher learning rates can speed up learning, they can also lead to oscillatory and unstable behavior, especially during collisions. The learning rates were empirically set to balance these factors.
- Future Work: The authors plan to implement the controller on a real robot arm and test its performance in pick-and-place tasks involving disturbances.
Summary updated