Periodic Reporting for period 4 - ROBOTGENSKILL (Generalizing human-demonstrated robot skills)
Okres sprawozdawczy: 2023-04-01 do 2024-03-31
To this end, Robotgenskill addressed the following research question: how can a robot generalize a skill that has been demonstrated in a particular situation and apply it to new situation? Robotgenskill focused on skills involving rigid objects manipulated by a robot or by a human, and followed a model-based approach consisting of: (1) extracting an invariant skill representation from the demonstration data; (2) creating a new instance of the skill for a new situation by solving an optimal control problem in which similarity with the invariant representation is maintained while complying with the constraints imposed by the new context. This approach stands in contrast with mainstream AI approaches, which are purely data-driven and hence do not use physics-based models.
The major breakthroughs achieved by Robotgenskill are that the required number of demonstrations and hence the training effort drastically decrease compared to mainstream AI approaches. Furthermore, similarity with the demonstration is maintained, preserving the humanlike behavior, while the task can easily be adapted to new situations without requiring much additional training effort.
As demonstrated during the project, the methodology developed in Robotgenskill can be applied to program advanced and sensor-based robot skills involving motion in free space (e.g. human-robot hand-over tasks) or skills involving motion in contact (e.g. assembly or cleaning). Such tasks are important in industrial and non-industrial settings such as health care and home environments. The project has also generated results that are useful in other areas. Some of the concepts and tools can be used in biomechanics to model human joints, such as the knee. In addition, the numerical algorithms and software developed to speed up the solution of optimal control problems can be widely used in the robotics and control community to enable real-time control of complex systems.
We used this methodology to generate shape-preserving and reactive robot trajectories in free space, towards new and even moving targets. Building upon this, we showed that virtual demonstrations can be generated and added to human demonstrations to extend the extrapolation capabilities of robot motion models that are learned from demonstrations. This approach allowed us to drastically reduce the required number of human demonstrations, hence reducing the cost of commissioning robot tasks, while improving the robustness, reliability, explainability and hence also safety of the robot application.
For robot skills involving motion in contact, we developed a generic methodology based on similar concepts and procedures to extract a task model, but now also considering interaction forces. The skill model enables adaptation of the task execution to deal with changes between the demonstration context and the new task context. We validated the method for seven different applications including assembly tasks, surface following tasks and manipulation of articulated objects such as doors and drawers.
As for the biomechanics applications, we designed an optimal method for estimating the average axis of a human joint (e.g. the knee axis), with improved sensitivity to noise and including a dispersion analysis of the instantaneous knee axis. We also designed a novel procedure for estimating the knee flexion angle based on functionally defined coordinate systems. This procedure eliminates the need for a time-consuming calibration of the markers used for motion capture, which is important for clinical practice.
An unplanned outcome was the new numerical solver for Optimal Control Problems, referred to as FATROP, which achieves the same level of robustness as state-of-the-art generic nonlinear solvers such as IPOPT, but with a much smaller computation time. Since FATROP can efficiently solve a broad class of nonlinear optimal control problems, it is expected to impact other domains beyond the applications considered in the Robotgenskill project, such as drones, self-driving cars and machines like gantries, compressors and drivetrains.
demonstrations and has very limited generalization capabilities, i.e. abilities to generalize the learned skill to 1) a new task, 2) a similar tasks but in a different setting, or 3) even exactly the same task which encounters one or more additional constraints during execution which were not present during the demonstrations.
In this project we developed an approach to learn an explainable model from the demonstration data. Such model consists of relations (i.e. ‘constraints’, formulated as expressions) between the data. Innovative is also that we use both motion data and interaction forces. To enhance the generalization capabilities of the model, we represent the motion data and interaction forces in a coordinate-invariant way. So, our models do not depend on the particular reference frames or reference points that have been chosen during the demonstrations or during the analysis. We also make sure that in our approach the number of hyper parameters in the models (the ‘magic numbers’) is limited to the absolute minimum. If there are hyper parameters, they have a physical meaning so that they can be determined as much as possible beforehand and do not have to be tuned experimentally for a given task and context.
Generalization of a skill is further supported by a constraint-based robot controller which tries to satisfy 1) the learned model constraints, 2) additional model constraints for the particular application, if any, and 3) additional constraints corresponding to the new context in which the skill is applied. The advantage of our approach is that it requires a lot less demonstrations and has much better generalization capabilities compared to a purely data driven approach.