NASA's Valkyrie robot is, arguably, one of the most sophisticated and potentially capable humanoids in existence. The key here is "potentially," because in addition to the very fancy and expensive hardware that was put into the robot at its conception, it has to have the software and controls that enable it to, you know, do stuff. Reliably.
Valkyrie didn't have that great of a time at the DARPA Robotics Challenge Trials, and we never quite figured out why. We certainly noticed that we didn't hear much about the robot after it didn't make the cut last December. So what's up with Val?
At the Robotics Science and Systems conference at UC Berkeley last month, we got an update on Valkyrie, including a description of what went wrong at the DRC and NASA's plans for the robot going forward.
The NASA JSC presentation at RSS was supposed to have been given by Valkyrie program manager Nic Radford, but he ended up not being able to make it at the last minute. Instead, Peter Neuhaus (from the IHMC DRC Team) filled in for him, which turned out to make a lot of sense, as IHMC assumes a key role in the Valkyrie project (more on that later).
We were at the presentation and took a bunch of notes on all the most interesting stuff, which is presented below in InfoNugget™ format:
• The design for Valkyrie started in October of 2012, and alpha assembly was finished in July of 2013. Robonaut 2 was an inspiration for Valkyrie, but the design was completely new. A team of 35 designers collaborated on Valkyrie, which was a challenge by itself. The video below shows some of that process:
This video shows various stages of the making of NASA's humanoid robot Valkyrie. The Valkyrie humanoid robot was built at NASA Johnson Space Center by a medium size group of of mechanical, electronic and embedded systems designers and with leadership from Nic Radford and Rob Ambrose. The Human Centered Robotics Lab at UT Austin provided expertise in the design of rotary series elastic actuators and inspiration for the design of the linear series elastic actuators on the robot's ankles and torso. In particular, the linear series elastic actuators were inspired by the UT-SEA design developed during Nick Paine's thesis as a high performance series elastic actuator. UT Austin also provided expertise on developing robust force controllers of the series elastic actuators and thorough testing on torque tracking and joint position impedance control.
• After July, a lot of time was spent on developing soft goods as opposed to software development, which the team admits was probably not the best way to focus their resources. But what really messed things up for the team was the federal government shutdown in October. The shutdown itself was 16 days, but by the time all of the contract people were back up to speed, it was a solid month out of development and testing.
• Valkyrie was "overloaded with sensors." Since the team didn't have a good idea (in advance) about what sensors would be necessary, they just crammed in everything that would fit. This includes three LIDARs, four HD cameras, and six depth map cameras, just in visual sensing. Also, more sensors means more redundancy if something stops working.
• At the DRC Trials, Valkyrie experienced a "networking issue" that prevented the team from scoring any points. In the garage before the DRC Trials began, everything worked fine. But on the course itself, the JSC Team "could not communicate with the robot at all." They would later discover the culprit: a network traffic shaping tool that they'd added to their code and that ended up blocking data from the operator to the robot. This manifested itself as a "major instability in the control system," preventing the robot from functioning almost completely.
• The first day was the "easy day" in terms of the tasks the robot would face. Valkyrie's networking issue was fixed in time for the last event of the day (opening the three doors), which meant the bulk of Valkyrie's opportunities for scoring points were missed. In practice (mock DRC Trials back at JSC), Valkyrie was reliably scoring 6-8 points, and if they'd had the hard tasks on Day 1 instead of Day 2, they would have identified the issue and been able to fix it in time for the easier tasks.
• Rather than continue developing Valkyrie on its own, NASA has used a National Robotics Initiative grant to fund a research program, "Toward Humanoid Avatar Robots for Co-exploration of Hazardous Environments," in collaboration with IHMC. "The goal for NASA is to have a humanoid robot that might assist in space exploration," and IHMC is working on the controls for that, with Valkyrie as a platform. The end goal is to have platform agnostic control software that can control either Valkyrie or ATLAS (and eventually other humanoid robots).
• NASA is hoping to use robots like Val for Mars exploration as a humanoid avatar. The robot will get to Mars early, before any humans, and then be "assistively teleoperated" into assembling structures and exploring so that everything is prepared for the eventual arrival of some weak and fragile humans. And while the humans are there, the robots will perform maintenance and repairs to free those humans to do more interesting things. NASA also wants to be able to use humanoid robots a little closer to home, for search and rescue and disaster response, bringing things neatly back to the DRC.
• As of a few weeks ago, Valkyrie was at the "basic walking" stage, as NASA JSC and IHMC work to improve the robot's control code. IHMC had also completed an assessment of Valkyrie and proposed a series of next steps. We transcribe the highlights below:
IHMC Preliminary Assessment of Valkyrie Platform:
- Force control is really good.
- It seems like we can also achieve fairly stiff position control, but we need to evaluate this more.
- Joint torques and speeds seem good so far. Might need a little more knee torque (the ability to do a deep knee bend).
- Need a little more range of motion in the hip joints.
- Need a larger field of view LIDAR in the head.
- Need an easier way to calibrate the joint position and torque sensors.
- Ability to disassemble the robot for shipping and maintenance is great.
- Robot is fairly easy to access internals for maintenance.
- Robot is pretty repeatable and reliable, but needs extensive cycle testing to determine and fix failure modes.
- Nice safety systems for human interaction; we feel comfortable working with the robot (unlike ATLAS).
- Finish tuning existing controls for Valkyrie and demonstrate on DRC cinder block walking, doors, valve, and drill tasks.
- Continue on subtasks toward Mars co-exploration.
- Manipulate more objects.
- Maneuver through jungle-gym like structure.
- Push recovery.
- Walking over rocks with partial foot contact.
- Standing up from prone?
- Continue working with JSC
- Feedback on Valkyrie hardware.
- Integrate walking algorithm with JSC high level software and user interface.
If everything goes well, Valkyrie will be a possible Track D (unfunded) entry in the DRC Finals, potentially as a joint team with IHMC and NASA JSC. We're very excited to see this happen.
Lastly, we've heard that some of the team that was working at JSC to develop Valkyrie as since moved on to found their own startup. But that deserves its own article, so we'll have more for you in a week or two.