Teaching Robots to Interact with Humans in a Smart Environment

1 Introduction

As estimated by Alzheimer's association in 2017, number of individuals more than 85 years of age is projected to triple by the year 2050. Currently, 50% of older adults require assistance with daily activities of living and the annual cost of this assistance is close to $2 trillion [Ernst and Hay, 1994]. Thus, alternate methods of healthcare are needed to preserve their quality of life. With the advancement in technology, there has been a gradual shift of focus on creating robots to provide assistance in healthcare. The fields of machine learning and psychology are in pursuit of helping elderly adults age sustainably at home [Gessl et al., 2019]. In recent years, with advancement in the field of robotics and reinforcement learning there has been an active research in the direction of creating more efficient robots and subsequently leading to better human robot interaction, a research field dedicated to designing, evaluating and understanding robotics systems to be able to efficiently communicate with people. In one of the recent works by Wilson et. al., a multi agent robotic system was developed that assists elder adults in a smart environment [Wilson et al., 2019]. In this work, the robot is in conjunction with a smart home agent. The task of the smart home agent is to track user activities and signal the robot to offer assistance if a person needs one. However, this work currently relies on Simultaneous Localization and Mapping (SLAM) for navigation. The basic problem for SLAM is that the map of the environment should be completely known. Instead, a more robust and intelligent navigation system should be in place so that the robot can provide better assistance. In this research we propose a map less navigation method to successfully locate the human in the smart home. The goal is to generalize the navigation problem with the primary goal of locating the human. The primary construct of the idea is to leverage the information from the LIDAR sensors and visual signals (camera images) to assist in predicting the movement of the robot in the smart environment. The final goal of the robot will be to reach out to the person to offer assistance. The information from the LIDAR sensor is same everywhere, so with this technique, there will be no need to learn the map of a new environment. Thus, it will result into learning more generalized policy for elder care assistance and can be applied to every smart home environment. This research envisions creating an end to end algorithm capable of successfully navigating to human and then being able to interact with them in a complementary pose.

2 Research Problem

2.1 Problem Statement

The goal is to perform navigation in an unknown environment towards a potential human target present in the environment. We propose the use of LIDAR sensor information in the state space. In order to capture information of human pose we run a pose detection algorithm on each image obtained from the camera mounted on the robot and append the number of human poses detected in the state space. The design of reward function remains a crucial step in this setup. Reward needs to be tied to the movement of the robot as well as the number of humans detected in each frame. The goal of the robot is to find humans and if it successfully finds humans then its goal is to approach them.

2.2 ProblemFormulation

Inspired by the work of [Dobrevski and Skocaj, 2018] we define the state space by a vector st = [l1...l30, vt], where l1 to l30 correspond to the LIDAR readings obtained and vt corresponds to the number of human poses detected in each frame. The existing approaches for mapless navigation have information about the goal, the distance to the goal, orientation of the goal or in cases of target driven visual navigation the target image is provided[Zhu et al., 2017; Dobrevski and Skocaj, 2018; Lathuili`ere et al., 2019]. In our case the problem is much more difficult, the goal's location and pose is unknown. However, goal is always a human in our case. We propose to use realtime multi person 2D pose estimation which is shown to perform pose estimation at a speed of 30 frames per second [Cao et al., 2017]. The action space is defined by at = [⋄,←,→, ↑, ↓], which denote stay, turn left, turn right and go back respectively. We plan on adding a continuous action space of the pan tilt angle which will help the robot complement the gaze of the human. Reward function is the most important formulation of the Markov Decision Process (MDP). In this case the reward should be tied to the number of human pose detected since the main goal of the robot is to find the human. Additionally, the robot's local goal is to avoid obstacles on the path. Therefore, we define a reward function as Rt = γNt +dt +rt, where Nt is the number of human pose detected, dt will be the proximity to the nearby objects (measured by the depth sensor on the robot), and rt is defined as high negative reward for any collision and low negative reward for each action taken at a time step. The low negative reward will aid in finding the shortest possible path to reach human. In RL, model parameters are learned on sequences of state, action and rewards, termed as episodes. The goal is to learn a policy that maps the state to actions. Qfunction (action-value function) is the expected future return from a state S, taking action At and then following a policy π. Learning a best policy corresponds to optimization problem. To solve this optimization problem, Q-function is modelled by a neural network which takes as input the state space St and outputs corresponding actions pertaining to the best policy. As inspired by the works of [Lathuili`ere et al., 2019], Long short term memory networks (LSTM) [Hochreiter and Schmidhuber, 1997] will be used. LSTM being a recurrent neural network has shown to exhibit dynamic behavior for temporal sequences.

3 Research Plan

As discussed earlier, we already have a robot setup in a smart home environment [Wilson et al., 2019]. However, a new learning system needs to be developed and deployed. A simulation will be programmed using openAI gym and Gazebo [Zamora et al., 2016]. Algorithm will be trained on this simulator and then transfer learning will be applied to transfer knowledge to a real robot [Taylor and Stone, 2009]. If everything is successfully learned, we plan to deploy the learning system in smart robot in smart environments and perform a study with older adults to research the advocacy of robot assistance for them.

4 Conclusions and FutureWork

We propose a method to perform a mapless navigation using reinforcement learning, and envision to deploy this learning system in a real robot. This approach relies only on the sensor information and the robot learns a policy using raw features. Once the approach is successful we plan to extend the state space with audio signals, which would help the robot to search a person more efficiently and also aid the robot to gaze at the person.

Acknowledgements

The author would like to thank Dr. Matthew E. Taylor and Dr. Diane J. Cook for their guidance and support.