TAPIR: Toolkit for approximating and Adapting POMDP solutions In Real time


TAPIR is a C++ implementation of the Adaptive Belief Tree (ABT) algorithm. ABT is an online and anytime approximate POMDP solver capable of computing approximately optimal solution of various robotics problems, including robots operating in partially known and dynamic environments, in real-time.

TAPIR automatically updates the POMDP model as needed when the environment or the robot's understanding of its environment changes. It then adapts the POMDP solution to modifications of the POMDP model without the need to reconstruct the policy from scratch.

TAPIR provides a command-line interface, as well as an interface with ROS and V-REP.

For bug reports and other feedback, please e-mail us at rdl.algorithm@itee.uq.edu.au.

Videos (TAPIR with ROS+V-REP Interface)


A video demonstration of TAPIR. This video consists of two segments.
First is a Target Tracking scenario in continuous state space when the environment map is not known a priori. During runtime, the robot scans its environment, builds a map, updates the POMDP model, and adjusts the POMDP policy it uses.
Second is Tag benchmark in larger environment, showing how TAPIR adjusts its policy in response to changes in the environment at runtime.

Downloads

Release 0.3.1.c (20 March 2017)

TAPIR v0.3.1.c
Quick Guides: To get started | To implement a new POMDP model | Using TAPIR with ROS and V-REP
Documentation: Online | Download
What's new?

  • (0.3.1.c) Minor fixed for compatibility with gcc 5.4 and above.
  • (0.3.1.b) Minor fixed on the default configuration of the problems.
  • (0.3.1) Fixed compilation for ROS (added missing GPS-ABT source files to CMakeLists.txt).
  • Added GPS-ABT for continuous action spaces.
  • Added a ”–no-load” option to ./simulate so that it can be run without a need for serializing policies.

Release 0.2 (1 December 2014)

TAPIR v0.2
Quick Guides: To get started | To implement a new POMDP model | Using TAPIR with ROS and V-REP
Documentation: Online | Download
What's new? Better interface for ROS and V-REP.

Release 0.1 (13 July 2014)

Publications

  • . . . . . . Vol. . No. . pp. . ed. . ()

  • People

    • Hanna Kurniawati (research group leader)
    • Dimitri Klimenko (developed the core TAPIR library)
    • Joshua Mun Song (developed the ROS + V-REP interface)
    • Konstantin Seiler (adapted TAPIR for continuous action spaces)
    • Vinay Yadav (helped to develop the basic ABT algorithm)