TL;DR
This paper introduces a novel approach to controlling redundant robots by combining learned forward kinematics models with operational space control. This allows robots to perform multiple hierarchical tasks simultaneously, even when those tasks are not perfectly compatible, by leveraging learned null space projectors. The method demonstrates better versatility than existing approaches that learn inverse kinematics directly.
What Problem Does This Paper Address?
Modern robots are increasingly deployed in complex, unstructured environments, requiring them to handle a multitude of tasks and adapt to uncertainties. Redundant robots, those with more degrees of freedom than strictly necessary for a given task, offer greater flexibility but introduce challenges in control. Specifically, controlling these robots to perform multiple tasks hierarchically, where one task might depend on another, is difficult. Traditional methods often struggle with learning inverse kinematics directly, especially in redundant scenarios where infinite solutions exist, leading to potential information loss and singularity issues. This paper tackles the problem of enabling redundant robots to execute complex, hierarchical tasks by learning their forward kinematics and using this learned model within the operational space control framework.
Methods
The core of the proposed method lies in learning the forward kinematics model of the robot using Locally Weighted Projection Regression (LWPR). LWPR is chosen for its ability to provide accurate approximations in high-dimensional spaces and its efficient computation of the Jacobian matrix as a byproduct of learning the forward model.
Instead of directly learning the inverse kinematics (which can be problematic for redundant robots), the authors learn the forward mapping from joint space to task space. This learned forward model is then used to derive the Jacobian matrix at the velocity level. With the Jacobian, they can then compute the necessary inverse kinematics and null space projectors required for operational space control.
The paper compares two redundancy resolution schemes:
- Projection Method: This method uses null space projectors to handle secondary tasks or constraints without perturbing the primary task. It's particularly effective for hierarchical task execution.
- Extended Jacobian Approach: This method attempts to combine multiple tasks into a single, larger Jacobian matrix. The authors adapt this approach by learning the forward kinematics and then inverting the extended Jacobian, rather than directly learning the inverse kinematics as in prior work.
Datasets
The paper does not use pre-existing datasets. Instead, it relies on a simulated 3-degree-of-freedom planar robot. The "dataset" for training the LWPR model is generated through a "motor babbling" phase. This involves randomly moving the robot's joints and recording the corresponding task space outputs. This process is repeated for a specified number of samples (e.g., 2000 or 5000) to initialize and adapt the LWPR model. The evaluation is performed using this simulated robot under various task scenarios.
Key Insights and Ideas
- Learning Forward Models is Key for Redundancy: Directly learning inverse kinematics for redundant robots can lead to a loss of information about the robot's redundancy. Learning the forward model, however, preserves this information, allowing for more flexible and robust control.
- LWPR for Jacobian Derivation: LWPR is not only effective for learning the forward kinematics but also provides the Jacobian matrix "for free," which is crucial for operational space control and redundancy resolution.
- Hierarchical Task Execution with Learned Projectors: The ability to learn null space projectors from learned forward models enables the hierarchical execution of multiple tasks, even when they are not perfectly compatible. This is a significant step towards more complex robotic behaviors.
- Projection Method's Versatility: The projection method proves to be more versatile than the extended Jacobian approach, especially when dealing with learned models and incompatible tasks. It allows for separate learning of tasks, making the system more modular and adaptable.
- On-line Adaptation is Crucial: The experiments highlight the importance of on-line model adaptation during control. Even with initial learning, the robot's performance improves as it continues to learn while executing tasks, compensating for model uncertainties and errors.
Key Results
- Under-Constrained Case: The learned model, after initial babbling, allows the robot to achieve reaching tasks with a prescribed precision (0.01m). The end-effector trajectory converges to a smooth, expected path. Both the projection and extended Jacobian methods perform comparably.
- Fully Constrained Case: The system successfully executes two tasks simultaneously, with the projection method and extended Jacobian approach showing similar precision. Errors are managed effectively due to on-line model adaptation.
- Over-Constrained Case: The projection method demonstrates its effectiveness in handling incompatible tasks. The primary task is achieved with high precision, while the secondary task is executed with the minimum possible error without disturbing the first task. This is a significant result for learned models in over-constrained scenarios.
- NMSE Reduction: The Normalized Mean Square Error (NMSE) for predicting task space velocity decreases significantly during the babbling phase, indicating effective learning. A babbling phase of 5000 samples is shown to be sufficient for good prediction and Jacobian estimation.
Limitations and Future Work
- Babbling Phase Necessity: The paper emphasizes the necessity of a babbling phase for initializing the LWPR model. The number of receptive fields required can be substantial, influenced by the desired precision and the need to cover the joint space adequately.
- Error Propagation: Errors in Jacobian prediction can lead to disturbances in task achievement, particularly in the projection method when computing projectors.
- Future Work:
- Extending the framework to trajectory tracking tasks, which would likely require more on-line learning during control.
- Investigating the behavior under perturbations to further validate on-line adaptation capabilities.
- Exploring different types of inversions and task combinations, including joint limits avoidance and external collision avoidance.
- Analyzing the scaling effects of increasing the number of degrees of freedom on learning complexity.
Summary updated