DARPA challenges virtual robots before the real ones step in
Teams vying for a spot in the historic DARPA Robotics Challenge (DRC) must first prove themselves in the Virtual Robotics Challenge (VRC) later this month. The VRC digitally simulates the physical challenges slated to take place at the tail end of the year, where real robots will get down and dirty for the first time. A total of 26 teams qualified to take part in the VRC, but only eight of them will earn the privilege of working with their very own ATLAS humanoid. Others will participate with their own unique robots.
The challenge tasks are based on past disasters and the input of veteran firefighters, nuclear engineers, and other first responders who were asked what a disaster response robot should be able to do. The whole point is to create a robot that can stand in for people in extremely dangerous or deadly environments, such as during a melt down at a nuclear plant.
The simulation itself runs in Gazebo and is being designed by the Open Source Robotics Foundation (OSRF). The idea is to replicate the physics and challenge restrictions as closely as possible, ensuring that teams won't be in for any nasty surprises when they port their code to the real robot.
For example, the real robot's limbs tend to oscillate a bit before coming to a stop, so the simulated robot exhibits the same behavior. And the teams won't benefit from a floating third-person perspective as they would playing a video game; they'll have to operate the robot only through its on board cameras.
It may not look all that pretty, but that's because the focus is on realistic physics simulation rather than graphics. You can see what the simulation looks like in the following video, which gives an overview of the included challenge tasks:
The VRC isn't going to be broadcast live, but the OSRF says the simulations will be recorded and made public. Activities will include picking up objects, turning valves, and driving a vehicle to a destination. However, each team will be presented with a slightly randomized version of each task, ensuring their control software will have to deal with a diverse range of parameters.