Trajectory prediction in cluttered voxel environments

Abstract

Trajectory planning and optimization is a fundamental problem in articulated robotics. It is often viewed as a two phase problem of initial feasible path planning around obstacles and subsequent optimization of a trajectory satisfying dynamical constraints. There are many methods that can generate good movements when given enough time, but planning for high-dimensional robot configuration spaces in realistic environments with many objects in real time remains challenging. This work presents a novel way for faster movement planning in such environments by predicting good path initializations. We build on our previous work on trajectory prediction by adapting it to environments modeled with voxel grids and defining a frame invariant prototype trajectory space. The constructed representations can generalize to a wide range of situations, allowing to predict good movement trajectories and speed up convergence of robot motion planning. An empirical comparison of the effect on planning movements with a combination of different trajectory initializations and local planners is presented and tested on a Schunk arm manipulation platform with laser sensors in simulation and hardware.

A robot path planning framework that learns from experience

“…Our approach differs in that we are concerned with selecting a single path from the library to bring the robot from the start to the goal. Similar to our approach, Jetchev and Toussaint [16] use a library of paths to aid in computing a new path in a new environment. However, their paths are computed in endeffector space, whereas ours reside in the higher-dimensional C-space.…”

STOMP: Stochastic trajectory optimization for motion planning

“…A possibility for future work is to augment this local trajectory optimizer with a trajectory library approach, which can recall previous trajectories used in similar situations, and use them as a starting point for futher optimization [20]. The STOMP algorithm could also be applied to problems in trajectory-based reinforcement learning, where costs can only be measured by execution on a real system; we intend to explore these avenues in future work.…”

Learning from Experience in Manipulation Planning: Setting the Right Goals

“…Previous work [1,3] proposed solving this problem by learning to index into a dataset of examples. This approach is limited by the dataset of previously executed trajectories, much like, for example, the earlier work in object recognition was limited by labeled images it used.…”

“…They focused on predicting globally optimal trajectories: given a training dataset of situations and their globally optimal trajectories, predict the globally optimal trajectory for a new situation. Much like Case-Based Reasoning, their approach predicted an index into the training dataset of trajectories as the candidate trajectory [1,2] or clustered the trajectories and predicted a cluster number [1,3]. Since prediction is not perfect, a post-processing stage, where a local optimizer is initialized from the prediction is used to converge to the closest local minimum.…”