Handling crowdsourced data using state space discretization for robot learning and synthesizing physical skills
- PDF / 1,944,476 Bytes
- 13 Pages / 595.276 x 790.866 pts Page_size
- 18 Downloads / 171 Views
REGULAR PAPER
Handling crowdsourced data using state space discretization for robot learning and synthesizing physical skills Leidi Zhao1 · Lu Lu1 · Cong Wang1 Received: 14 July 2020 / Accepted: 19 October 2020 / Published online: 12 November 2020 © Springer Nature Singapore Pte Ltd. 2020
Abstract Intelligent physical skills are a fundamental element needed by robots to interact with the real world. Instead of learning from individual sources in single cases, continuous robot learning from crowdsourced mentors over long terms provides a practical path towards realizing ubiquitous robot physical intelligence. The mentors can be human drivers that teleoperate robots when their intelligence is not yet enough for acting autonomously. A large amount of sensorimotor data can be obtained constantly from a group of teleoperators, and processed by machine learning to continuously generate and improve the autonomous physical skills of robots. This paper presents a learning method that utilizes state space discretization to sustainably manage constantly collected data and synthesize autonomous robot skills. Two types of state space discretization have been proposed. Their advantages and limits are examined and compared. Simulation and physical tests of two object manipulation challenges are conducted to examine the proposed learning method. The capability of handling system uncertainty, sustainably managing high-dimensional state spaces, as well as synthesizing new skills or ones that have only been partly demonstrated are validated. The work is expected to provide a long-term and big-scale measure to produce advanced robot physical intelligence. Keywords Robot learning · Physical intelligence · Robotic manipulation · Crowdsourcing
1 Introduction Physical intelligence such as dexterous manipulation and dynamic mobility is crucial for robots to interact with the real world. Methods for realizing robot physical intelligence generally sit in two types. One relies on motion planning based on analytical models derived from laws of physics. Representative work includes the dynamic re-grasping robot hand by the University of Tokyo Furukawa et al. (2006) and the bipedal robot Cassie by Agility Robotics Ackerman (2017). Despite the rigorous guarantee of stability and the full utilization of mechanical potentials, the extensive * Cong Wang [email protected] Leidi Zhao [email protected] Lu Lu [email protected] 1
New Jersey Institute of Technology, Electrical and Computer Engineering, 323 Martin Luther King Blvd, Newark, NJ 07102, USA
13
Vol:.(1234567890)
case-specific engineering and complex ad hoc analytical models required by such methods hamper their ubiquity. Thanks to the advance of artificial intelligence, robot learning methods such as learning from demonstration (LfD) Argall et al. (2009) and reinforcement learning (RL) Sutton and Barto (2018) have successfully allowed robots to acquire physical skills with less reliance on analytical models. Most of such methods adopt a “policy search” framework Deisenroth et al. (2013) and
Data Loading...
