Example-driven trajectory learner for robots under structured static environment

Abstract

With the breakthroughs in machine learning and computing infrastructures that have led to significant performance improvements in cognitive robotics, the challenge of continuous-trajectory task creation persists. This challenge stems from the need to account for inter-joint relationships, which define constraints between different robot joints due to the kinematic structure, and intra-joint relationships, which are constraints within a single joint like limits. Accounting for these coupled, nonlinear inter-joint and intra-joint relationships is crucial for trajectory planning. However, various constraints in the physical capability of robots, environmental changes, and long-time reliance on sequential dependencies between these inter-joint and intra-joint relationships make the work of modifying robot trajectories exceptionally hard. Many robot environments function under structured static work-cell completing extended series of subtasks. The conventional descriptors for robot trajectory rely on symbolic rules with human intelligence, which involves skilled individuals and possesses significant limitations, such as being time-consuming and exhibiting low flexibility even for minor changes, due to the static nature of task descriptions alone. The suggested technique employs a probabilistic network and data-efficient modelling termed generative adversarial networks, which learns the underlying constraints, probability distributions and arbitrations, along with generating trajectory instances at each time of sampling. Integrating prior knowledge into the symbolic trajectory learner as a dataset facilitates the learning process. The model assessment was carried out by utilising a custom-built dataset in a simulation based environment. This research also proposed two GAN inversion methods to compute the generated trajectory and its closest instance in the dataset. Furthermore, GAN Inversion method I and II calculated the robot path accuracy in extrinsic generative models yielded path position accuracy of 9.2 cm and 4.9 cm respectively. In addition to that, the study contributes a probabilistic model for interpolating between various trajectories to generate new trajectories.