Imitation learning approaches achieve good generalization within the range of the training data, but tend to generate unpredictable motions when querying outside this range. We present a novel approach to imitation learning with enhanced extrapolation capabilities that exploits the so-called Equation Learner Network (EQLN). Unlike conventional approaches, EQLNs use supervised learning to fit a set of analytical expressions that allows them to extrapolate beyond the range of the training data.
Learning and extrapolation of robotic skills using task-parameterized equation learner networks
We augment the task demonstrations with a set of task-dependent parameters representing spatial properties of each motion and use them to train the EQLN. At run time, the features are used to query the Task-Parameterized Equation Learner Network (TP-EQLN) and generate the corresponding robot trajectory. The set of features encodes kinematic constraints of the task such as desired height or a final point to reach. We validate the results of our approach on manipulation tasks where it is important to preserve the shape of the motion in the extrapolation domain. Our approach is also compared with existing state-of-the-art approaches, in simulation and in real setups. The experimental results show that TP-EQLN can respect the constraints of the trajectory encoded in the feature parameters, even in the extrapolation domain, while preserving the overall shape of the trajectory provided in the demonstrations.
Trajectory Adaptation from Demonstrations with Constrained Optimization
This paper proposes an approach for the adaptation of robot trajectories taken from a set of demonstrations.
The problem is formulated as a constrained optimization problem where the set of demonstrations are used as target values to build a Quadratic Program (QP). The constraints constitute the adaptation’s conditions of the new trajectory, e.g. new initial or final points or keep the trajectory within a specific range. The performance of our approach is verified in the adaptation of a set of demonstrations taken from a Pandarobot for new conditions
Best Poster Award
Austrian Robotics Workshop 2022
Multi-vehicle coordination based on hierarchical quadratic programming
This paper presents an optimization-based control scheme for generating online multi-vehicle coordination behaviors to accomplish missions in indoor environments. The proposed control scheme relies on the use of hierarchical task functions in terms of the multi-vehicle configuration variables. The task functions are related to individual and group obstacle avoidance, reaching fixed targets, group trajectory tracking, maintaining formations, enclosing the group within a geometric area, among others. The stack of hierarchical tasks
automatically handles possible conflicts between them. Quadratic programs are formulated for explicitly solving inequality and equality task constraints at any hierarchy. In addition, a finite state machine is employed to build complex group behaviors for successfully fulfilling group missions. The proposed control scheme is demonstrated on two experiments with static and moving obstacles where a group composed by six vehicles tracks a predefined trajectory for the center of mass of the group. In the second experiment, the group is asked to clean the workspace by pushing movable objects.
Control Engineering Practice, 2019
Return
Research projects
Multi-vehicle coordination based on hierarchical quadratic programming
This paper presents an optimization-based control scheme for generating online multi-vehicle coordination behaviors to accomplish missions in indoor environments. The proposed control scheme relies on the use of hierarchical task functions in terms of the multi-vehicle configuration variables. The task functions are related to individual and group obstacle avoidance, reaching fixed targets, group trajectory tracking, maintaining formations, enclosing the group within a geometric area, among others. The stack of hierarchical tasks
automatically handles possible conflicts between them. Quadratic programs are formulated for explicitly solving inequality and equality task constraints at any hierarchy. In addition, a finite state machine is employed to build complex group behaviors for successfully fulfilling group missions. The proposed control scheme is demonstrated on two experiments with static and moving obstacles where a group composed by six vehicles tracks a predefined trajectory for the center of mass of the group. In the second experiment, the group is asked to clean the workspace by pushing movable objects.
Control Engineering Practice, 2019
Multi-vehicle coordination based on hierarchical quadratic programming
This paper presents an optimization-based control scheme for generating online multi-vehicle coordination behaviors to accomplish missions in indoor environments. The proposed control scheme relies on the use of hierarchical task functions in terms of the multi-vehicle configuration variables. The task functions are related to individual and group obstacle avoidance, reaching fixed targets, group trajectory tracking, maintaining formations, enclosing the group within a geometric area, among others. The stack of hierarchical tasks
automatically handles possible conflicts between them. Quadratic programs are formulated for explicitly solving inequality and equality task constraints at any hierarchy. In addition, a finite state machine is employed to build complex group behaviors for successfully fulfilling group missions. The proposed control scheme is demonstrated on two experiments with static and moving obstacles where a group composed by six vehicles tracks a predefined trajectory for the center of mass of the group. In the second experiment, the group is asked to clean the workspace by pushing movable objects.
Control Engineering Practice, 2019
Time convergence for multi-robot control tasks based on hierarchical quadratic programming
In this work, we developed a framework focused on define finite-time convergence for more than one task in the same stack of task for the multi-robot coordination control. Our approach allows to avoid conflict tasks during the the performance by solving the tasks in a hierarchical way via quadratic programming.
The main contribution of this work is to define a finite-time convergence independently for each task, regardless the type of task, type of constraint or its hierarchy.
Return
Experimental set-up using a camera on ceiling and vision algorithms with OpenCV librearies to implement cooridnation control strategies for group of mobile robots
The experimental setup was focused on implement and test coordination control strategies in multi-robot systems. The experimental setup consists of a camera on the ceiling, which works with 30 frames per second. Each robot has a distinctive mark on the top which helps the algorithm classify the robots each other.
A nonlinear control was implemented for performing synchronized tracking trajectories. The nonlinear control has two types of gains, tracking gains in charge of minimize the tracking error and mutual coupling gains in charge of generates a bilateral virtual topology connection between the robots.
When an external disturbance is presented in one or more robots, mutual coupling gains are in charge of keeping the formation as best possible way. There is a trade-off between the coupling gains and tracking gains. The control helps to perform the tracking trajectory whereas the formation structure is kept.
Tracking gains
Mutual coupling gains
Linear and angular control velocities
Tracking errors
For more details about the control
Return
Control considering
a dealy in the
state reception
Reduction of delay effect in the synchronization of two multi-robot remote systems using adaptive gains based on the Speed-Gradient method
The project was focused on reducing the delay effect in the synchronization of two multi-robots remote systems located in two different places. Each group of robots had to track a desired trajectory in their corresponding place and in a synchronized way. The group or robots have a state-feedback each other. During this exchanged information is when the delay-time come out.
A delay-time in the communication generates oscillations in the tracking trajectories. The larger is the delay-time, the bigger are the amplitude of the oscillations. If the delay-time is too large, the system will became unstable and the synchronization will break.
A nonlinear control was implemented for the group-tracking-trajectories. This nonlinear control has two types of gains, tracking gains; in charge of reducing the tracking error, and mutual coupling gains; in charge of generates a bidirectional virtual topology connection between the robots. This project was focused specifically on substitute the tracking gains which had been considered constant so far, by adaptive gains, which are a minimized function that depends on the delayed-state.
The experiment results show that the adaptive gains attenuate the effects of the delay-time. The amplitudes of the oscillations are significantly reduced, allowing to maintain the formation structure of the group of robots, even in the presence of external disturbances. Additionally it the method provides favorable results for variable-delay-times.
Tracking error vector with delay
Adaptive tracking gains
Speed-gradient method ensure the limit of Q tends to cero when t goes to infinite.
Constant gains
Robots in México
Merged robots representation
Robots in Netherlads
Adaptive gains
Robots in México
Merged robots representation
Robots in Netherlads
Average delay time throughout the experiment= 281 ms.
Available only in spanish version
Return
This brief proposes a hierarchical control scheme based on the definition of a set of multirobot task functions. To deal with the inherent conflicts between tasks, a strict hierarchy is imposed on them. We present a novel scheme that copes with two main difficulties shared in standard task-based controllers: 1) to impose a desired time convergence of tasks and 2) to avoid discontinuous task transitions occurred when a task is inserted or removed in the hierarchical structure. As a result, continuous input references are generated for the low-level control of the group. The validation is achieved in simulation and by performing an experiment with wheeled mobile robots.
Hierarchical task-based control of multi-robot Systems with terminal attractors
Paper
Return
Paper (available only in spanish version)
Feasible trajectories generation for the formation of group of mobile robots and obstacle avoidance using PRM and Dijkstra algorithms
In this project, the group of mobile robots have to go from an initial configuration to a final configuration, keeping a specific formation and performs obstacle avoidance. For this, the formation was model as a planar robot configuration with 4 DOF. We use the PRM algorithm for the configuration space sampling and the Dijkstra algorithm to get the shorter path between the initial configuration and the final configuration. A nonlinear control with mutual coupling gains was implemented to ensure the formation structure along the tracking.
Return
Reduction of delay effect in the communication of a differential mobile robot using particles filter algorithm
In this work, the particles filter algorithm was implemented to compensate the negative effects of delay time generated by the communication systems in the tracking trajectories of a mobile robot.
It is assume that there is a known delay time in the reception robot state coming from the sensor (In this case, a camera on the ceiling) and it is also assumed that there is a known delay time in the sending of the control velocities to the robot.
Our approach choice the best estimation based on a probability distribution function given by the particle filter algorithm
There are three main stage in our approach:
-
First stage: The first one is a prediction of the robot state based on the kinematic model from a delayed observation, considering the sent control in last iteration, considering the sending delay time and the reception delay time as well.
-
Second stage: In the second stage, the particles filter algorithm gets a better fit of the observation obtained in the first stage. This is aimed to know which will be the state of the robot when the control arrive to him, considering the sending delay time.
-
Third stage: In the third stage, a nonlinear control calculate a predictive control law for the robot based on the final observation obtained in the second stage.
The approach works similarly to a Model predictive control approach, but in this case instead of minimize a cost function in a time window, our approach choice the best estimation based on a probability distribution function given by the particle filter algorithm.
Paper (available only in spanish version)
Robot performance without Particles filter algorithm
Robot performance with Particles filter algorithm
-
Sending delay time= 100 ms
-
Reception delay time= 100 ms
Return
The project consist in a flight simulator prototype. The user interface was developed with LabVIEW using an inertial sensor and quaternions approach. The simulator shows the artificial horizon and orientation indicators (roll, pitch, yaw). Actuators of UAV are controlled with a joystick
Development of a flight simulator and a physical interface for an UAV
Available only in spanish version
