Understanding humans to make more intelligent robots
Humans are able to satisfy multiple objectives in parallel while performing a particular task. Let us assume a simple task like the one of watering plants. While performing this task, humans are able to satisfy multiple objectives in parallel. Humans should decide how to move the watering can to irrigate all plants, avoid colliding against the plants, maintaining the stability to avoid falling or taking into consideration their joint limits.
In our work, recently published in the Robotics Science and Systems conference (R:SS 2021) titled “Composable Energy Policies for Reactive Motion Generation and Reinforcement Learning”, we studied the problem of modelling this skill composition to provide our robots with more dexterous capabilities or to better understand and predict human motion.
How do we model Skill Composition?
Modelling by skill composition is a very interesting approach to have sample efficient and intuitive robot controllers. Let us assume that we want to represent the robot motion to solve the pouring task we have introduced before. Trying to model directly a policy that is able to satisfy all the desired objectives is hard. However, skill composition allows us to represent the desired robot motion as a combination of a set of simple modules, each one trying to satisfy a particular objective. This modularity is a desired property for robotics, allowing us to represent complex robot behaviours as a combination of multiple goals. The question we need to answer is, how should we model each module and how should we combine them?
In our work, we consider the solution of the skill composition as an inference problem. We are looking for the distribution that satisfy all the objectives jointly. Each module is represented by a distribution. For example, the distribution representing robot’s stability will be a function that will provide high probability to those actions that maintain the robot stable, while low probability will make the robot fall. Then, if we combine all these probabilities, the composed probability will provide high probability to those actions that satisfy all the desired objectives, while low to any action that fails satisfying a particular objective. The skill composition can be thought as a probabilistic AND operation.
Skill Composition in SHAREWORK
SHAREWORK project benefits from the use of Skill Composition for Human motion prediction. Thanks to skill composition, we have modelled structured probabilistic models to predict human motion. These models would allow the robotic systems to safely plan the motions adapting to predicted movements.
Given some recorded human data, the models will learn to represent the modules the human motion is constituted of and then, once the model is learned, it will provide probabilistic future movements of the human.
About the author
Julen Urain
Robotics Ph.D Student
Julen Urain is a Ph.D student in the Intelligent Autonomous Systems department in Technical University Darmstadt(TUDA) advised by Jan Peters. He holds a master in Robotics and Control, with a Master thesis in Robot-Human Interaction. His research interests are in the interplay of Geometry, Optimisation and Deep Learning for reactive motion generation in robots.
