- The paper introduces a probabilistic framework that learns task constraints and priorities from demonstrations to guide robot manipulation.
- It employs a Dirichlet Process Gaussian Mixture Model and Gaussian Mixture Regression to adapt task constraints to new situations.
- Experimental results on dual-arm tasks demonstrate enhanced flexibility and reduced reproduction error compared to manual tuning.
Learning Context-Adaptive Task Constraints for Robotic Manipulation
This paper introduces a method for robots to learn how to perform manipulation tasks in different situations by observing demonstrations. The core idea is to automatically figure out the best way to control a robot, including what it should pay attention to and how important each aspect is, based on data from those demonstrations. This method uses a probabilistic model to understand how different situations (contexts) affect the way a task should be done, and then uses this understanding to adapt the robot's control strategy to new situations. The approach is evaluated on a dual-arm robot performing manipulation tasks.
Constraint-Based Control and Task Specification
Constraint-based control is a technique that allows robots with many degrees of freedom to perform multiple tasks simultaneously. It involves formulating each task as a constraint within an optimization problem, where the solution represents the best robot joint commands for accomplishing all tasks. Task prioritization is crucial to handle potential conflicts, using either strict or soft priorities. While constraint-based control is effective, it typically requires a human expert to define task constraints and priorities, which is time-consuming and situation-specific. The approach described in the paper aims to automate this process by deriving task constraints and priorities from data and adapting them to new situations.
Learning Adaptive Task Constraints
The proposed approach uses programming by demonstration to generate training data across multiple task variations, referred to as context changes. A Dirichlet Process Gaussian Mixture Model (DPGMM) is employed to model the joint distribution of context variables and task constraints. Gaussian Mixture Regression (GMR) is then used to reproduce task constraints and priorities in unseen contexts. The context is represented as a vector of real-valued or categorical variables (one-hot encoded). After recording demonstrations, the data is re-sampled, temporally aligned, and normalized. Rotations are represented using elements of the Lie algebra so(3) to avoid issues with Euler angles and quaternions. The joint distribution of context and motion variables, P(v,x,κ), is learned as a DPGMM. Task constraints and priorities are reproduced iteratively using GMR, with variance in user demonstrations used to estimate task priorities. High variability corresponds to low priority and vice versa. Generalization to unknown contexts is achieved by performing demonstrations under multiple task variations and optimizing the DPGMM's hyper-parameters using leave-one-out cross-validation.
Experimental Evaluation and Results
The approach was evaluated on three dual-arm manipulation tasks: rotating an object, collaborative transport, and assembly. The experiments were conducted on a dual-arm robot with KUKA iiwa arms and Robotiq grippers. The robot's base frame and end-effector frames were selected as task frames. The evaluation focused on the approach's ability to generalize to unseen situations. The model was trained with data from some contexts and tested on others. The results demonstrated that the trained model could generalize over object size and rotation direction in the rotate object task. In the collaboration task, the approach could generalize about different start poses. The assembly task results also showed the ability to generalize with respect to previously unknown start poses. Model performance was analyzed using the mean absolute error (MAE) between the mean of the demonstrations and the reproduced motion. The results showed that GMR requires a relatively low number of training samples to achieve good generalization capabilities. The capability of the approach to estimate suitable task priorities and adapt them over time, with respect to different constraint variables, and with respect to different contexts was evaluated. The estimated task priorities reflect the importance of the different constraints. A comparison of the reproduction error using estimated soft task priorities, manually selected soft task priorities, and estimated strict hierarchies shows that estimating soft task priorities results in a lower reproduction error and allows more flexibility for executing additional tasks.
Conclusion and Future Work
The combination of constraint-based control and imitation learning shows promise for improving the usability, applicability, and autonomy of complex robotic systems. The system exhibits some limitations, including the need for context labeling by a human expert and the design choice of task frame selection requiring expert knowledge. Future work includes classifying the current context from recorded data, determining whether a demonstration belongs to a known or unknown context, and selecting optimal task frames from user demonstrations. Applying the approach to more complex scenarios and robots is also planned. The authors also suggest adding an optimization step that improves the task weights with respect to a suitable criterion, like manipulability.