Anytime Task and Motion Policies for Stochastic Environments

Naman Shah, Deepak Kala Vasudevan, Kislay Kumar, Pranav Kamojhalla, Siddharth Srivastava

[Proc. IEEE International Conference on Robotics and Automation (ICRA), 2020]

ICRA 2020 paper
Extended version
Source code

Presentation Video

Videos of experiments

In all the experiments below, "solution" refers to a tree-structured proper policy for the abstract SSP with motion plans for all of the robot's movements.

Cluttered Table

Goal: Pick up the black object.

• Several objects are crushable, i.e. the Fetch can crush them while picking them up. If the object is crushed it needs to be placed in the recycling box.
• Several objects obstruct the target object, which needs to be replaced somewhere else to replace to reach the target object.
• There is no designated free area where obstructing objects can be stashed.

Our system computes a solution that picks and repositions several objects (placing the objects in the recycling box if they are crushed) before picking up the target object.

Building structures with Keva planks

Goal: Build the given Keva structure using ABB YuMi.

• The user can place a plank at one of the predefined locations. YuMi observes the environment and picks the plank from that location.

Our system computes the solution which handles all contingencies related to the placement of the Keva plank by the user.

Structure 1: Tower (height: 12)

Structure 2: Twisted Tower (height: 12)

Structure 3: 3π

Structure 4: Triple Tower (height: 8)