Robotic Neural Networks Predicting the Change in a System

Emily Hannigan

Robotic Manipulation and Mobility Lab

Columbia University

3/3/2023

11AM – 12PM CST

https://utsa.webex.com/utsa/j.php?MTID=mebdf9475a3c660cf65f2d72f41ab598d

Abstract:  Robots have the potential to aid humans from search and rescue work to space exploration, to in-home assistance for those with limited mobility. One of the main milestones required to achieve these ends is a robust ability to manipulate objects and locomote in cluttered and changing environments. To achieve this, a robot needs an internal model of the dynamics of itself and the world: given the current state of the world (the robot and the environment), how will any given sequence of actions change the state of the robot and its environment? This predictive model allows a controller to test out actions “in its head” before actually taking any actions. One of the most promising avenues for creating robust and accurate forward models is to use a neural network to approximate the dynamics of the system. However, the state of the art still falls short in a few areas. While it is easy to learn a data driven dynamics model to predict the change of a system over a short horizon, predicting farther into the future is still an open challenge. Additionally, the more complex a robotic system is, the more variables the model needs to keep track of and predict (we call this the state or observation of the system); the higher the dimensionality of the state, the more data and computation power we need to learn a model. This work investigates ways to improve the long term accuracy of data driven dynamics models using temporal data. It also evaluates methods of reducing state space dimensionality by learning this lower dimensional representation alongside a forward model using a method we dub “Context Training”. Throughout the development of these ideas, we also argue for improved evaluation metrics for learned forward models.