Narrow Results

We present a neural network that learns to control approach and avoidance behaviors in a mobile robot based on a form of animal learning known as operant conditioning. Learning, which requires no supervision, takes place as the robot moves around an environment cluttered with obstacles and light sources. The neural network requires no knowledge of the geometry of the robot or of the quality, number, or configuration of the robot's sensors. In this article we provide a detailed presentation of the model, and show our results with the Khepera and Pioneer 1 mobile robots.

In this research, a stable biped walking pattern is generated using reinforcement learning. The biped walking pattern is chosen as a simple third order polynomial. To complete it, four boundary conditions are needed. The initial position and velocity and the final position and velocity of the joint are selected as boundary conditions. In order to find the proper boundary condition value, a reinforcement learning algorithm is used. Also desired motion or posture can be achieved using the initial and final positions. The final velocity of the walking pattern is chosen as a learning parameter. To test the algorithm, a simulator that takes into consideration the whole model of the robot and the environment is developed. The algorithm is verified through a simulation.

Adaptive and cooperative control of arms and fingers for natural object reaching and grasping, without explicit 3D geometric pose information, is observed in humans. In this study, an image-based visual servoing controller, inspired by human grasping behavior, is proposed for an arm-gripper system. A large-scale dataset is constructed using Pybullet simulation, comprising paired images and arm-gripper control signals mimicking expert grasping behavior. Leveraging this dataset, a network is directly trained to derive a control policy that maps images to cooperative grasp control. Subsequently, the learned synergy grasping policy from the network is directly applied to a real robot with the same configuration. Experimental results demonstrate the effectiveness of the algorithm. Videos can be found at https://www.bilibili.com/video/BV1tg4y1b7Qe/.

Flexible-joint manipulators are frequently used for increased safety during human–robot collaboration and shared workspace tasks. However, joint flexibility significantly reduces the accuracy of motion, especially at high velocities and with inexpensive actuators. In this paper, we present a learning-based approach to identify the unknown dynamics of a flexible-joint manipulator and improve the trajectory tracking at high velocities. We propose a two-stage model which is composed of a one-step forward dynamics future predictor and an inverse dynamics estimator. The second part is based on linear time-invariant dynamical operators to approximate the feed-forward joint position and velocity commands. We train the model end-to-end on real-world data and evaluate it on the Baxter robot. Our experiments indicate that hard-wiring the one-step forward dynamics prediction into the model improves the performance, compared to the same model without prediction. We compare joint position, joint velocity and end-effector position tracking accuracy against the classical baseline controller and several simpler models with payloads of variable weight.