DARPA Robotics Challenge Track B Team Steel

Atkeson Group, Robotics Institute, Carnegie Mellon University

Team Meetings Tuesdays 7pm NSH 3305

If you are interested in joining our team, come to a team meeting Tuesdays 7pm NSH 3305 and/or send email to cga@cmu.edu.

CMU Class Spring 2013

I am teaching a CMU class 16-264 Humanoids this Spring that focuses on how to do the DARPA Robotics Challenge. Keep an eye on this web page for further information.

Help Wanted!

We are looking for people to help out in the areas of:

Vision/perception of obstacles and rough terrain.
User interfaces: How can an operator help the robot, without slowing it down? There is a lot of interest in supervisory control and predictive control. What is the best videogame interface you have seen? How about the control interfaces used in Robot Combat League?
Simulation: Particularly Gazebo-based simulation and ROS-based inter-process communication.
Tuning: There will be a lot of parameters to tune to optimize performance. We need to do this on a large scale.
Testing: We need to exhaustively test our entry, both at the component level and by simulating the full challenge.
GPU programming: We are interested in using GPUs to go faster.
Real time OS: Linux: Xenomai, Windows: ?
Software engineering: Keeping us organized, version control, etc.
Crowd sourcing: Helping us integrate contributions from others, etc.
Outreach: How can we get lots of students and others in on the fun?
Webmaster: Think you can generate better web pages than these? Go for it!
Competition analyst: What can you figure out about other team's strategies from googling them and the questions, answers, and discussions in various forums? Some teams put out YouTube videos.

We are looking for CMU students. We also are willing to work with people not at CMU, and to help others form Track C and D teams. Contact us (cga at cmu.edu) and we can try to work something out.

Donate!

Please help us support participation by students and others by sending us a donation.

Credit Card

You can use the CMU Donation Page to donate using a credit card (on the 2nd page put "Other" as the Designation. A new box Preferred Designation will appear. Type "Team Steel" in that box.)

Check

You can send a check payable to "CMU Robotics Institute" to:
Chris Atkeson/Team Steel
CMU Robotics Institute
5000 Forbes Avenue
Pittsburgh, PA, USA, 15213

Step 1: The Virtual Challenge (VRC)

The Virtual Robotics Challenge will be June 17-24, 2013.

A description of the VRC tasks

The tasks for the Virtual Robotics Challenge:

Climb into a utility vehicle, drive along a roadway at no greater than 16 kph (10 mph, 4.444m/s), and climb out of the utility vehicle.
Walk across progressively more difficult terrain, for example, progressing from parking lot to short grass to tall grass to tall grass on slope to ditch to rock field. In the earlier terrain, the GFE Platform balancing and walking behaviors will suffice. In the later terrain, DARPA expects perception and footstep planning will be needed.
Connect a hose to a spigot and open the spigot. This is purely a manipulation task, that is, the robot does not need to travel more than a few steps to the work site. The Hose Task shall be decomposed into three parts: mating the hose fitting to the faucet, connecting (threading) the hose fitting to the faucet by rotation, and opening the spigot enabling fluid flow.

DARPA Robotics Challenge Background Information

Left: The Track B robot.

Video of related BDI robot Pet-Proto.

DARPA Robotics Challenge Website

There are two DARPA Robotics Challenge groups at CMU: A Track A group led by Tony Stentz that will build a robot, and a Track B group (Team Steel) led by Chris Atkeson that will use a DARPA-provided robot. More information on the Atkeson Team Steel effort is provided below. The DARPA-provided robot is coming from Boston Dynamics. The PETMAN videos on the Boston Dynamics Youtube Channel are relevant. Check this channel for new humanoid videos.

If you would like to join Team Steel, send email to Chris Atkeson: cga at cmu.edu

If you would like to joint Stentz's effort, send email to Tony Stentz: axs at rec.ri.cmu.edu

The DARPA provided simulator will use Gazebo

Some prior work on ball juggling by Team Steel

Videos, Papers, and Powerpoint

Videos of our work are on the web pages of alumnus Ben Stephens.

Papers from our group are available from Chris Atkeson's web page.

Powerpoint and movies shown at Dynamic Walking 2012 relevant to DRC.

Some prior work on devil stick juggling by Team Steel

Team Steel Approaches

Learning

We are using robot learning in several ways.

Learning by being taught or programmed. Human teachers or programmers explicitly tell the robot what to do in various situations, or teach or program rules.
Learning by watching. The robot learns by watching humans or other robots do the task(s) or related tasks. Also known as learning from observation, learning from demonstration, and imitation learning. We will collect a lot of data of humans doing the challenge(s), and use that data to help our robot learn. This is similar to behavior libraries in videogames.
Learning by thinking or dreaming. The robot learns by planning or mentally practicing the task(s). Also known as caching plans. We will build a big library of successful plans. While doing a task, the most appropriate plan or set of plans is chosen and modified to fix the current situation (plan adaptation or refinement).
Learning by doing. The robot learns from practicing the task(s). Also known as learning from practice.

Optimal Control

We use optimal control and optimization as our major planning tool. An important research question is whether optimization using models will work in the messy real world.

A related research question is whether we can generate robot behavior that is robust. In our previous work, our robots worked great in our lab, but software running on identical robots in other places did not work well. What will it take to write robot programs that work well in our lab, other lab situations, and in the messy real world?

A particular specialty of our group is fast robust policy (control law) optimization using multiple models (alternative universes). This approach generates behavior that works well for a variety of possible robots and worlds.

Be Human-Like

We will attempt to mimic human task strategies, and also the soft (compliant) human touch. Most robots today are very stiff, and it is difficult for the robot to let the task or environment guide movement.

Open Source/Crowd Sourcing and Outreach

We are very interested in figuring out how to get other people to help us, either by helping test, tune, or train the robot, or by writing programs we use. Can a disparate group succeed in this challenge, and generate high performance robot control? Or is design/programming by committee doomed to failure?

We also see this challenge as an opportunity to get students interested in robotics, engineering, and science. We will explore how we can facilitate STEM (Science, Technology, Engineering, and Mathematics) outreach.