Run, Robot, Run! Our First Field Test on Forest Terrain

Our team recently took our firefighting robot out to Catheys Valley, a town about 35 minutes from UC Merced, to see how it would perform on mountainous terrain. Equipped with technology that allows it to lock onto and follow a target person, the robot is designed to literally lighten the load for forestry workers by carrying wood waste as it trails behind them. Previously, we had only operated the robot on flat concrete without a wagon attached. Before deploying it in the real world, we need to prove that our robot can efficiently haul wood on uneven forest ground—the reason for this field trip! 

We started by testing the follow-me technology in the new landscape full of grass, trees, and rocks, all of which could potentially disrupt tracking. At first, the robot would intermittently stop following the target person due to some untested changes made to the code. Luckily, our technical leads, Rodolfo and Kevin, fixed the issue by using the “git diff” command, used for comparing and editing two versions of the same file, to change the code back to its original version. Kevin explained, “I'm not sure what the problem was, but the workflow of frequently ‘version controlling’ the software and being able to revert back to known working versions is a very critical process for the engineering work. It's a very useful debugging tool and skill for software deployment and testing.” After getting it to work normally, the robot was able to trail smoothly behind its target on the hilly terrain—a big win for the team!

Figure 1. Our firefighting robot in the field!

That said, Kevin raised a valid concern: very tall grass could interfere with the robot’s ability to detect a full human skeleton, which is required for accurate tracking. The target person’s feet, ankles, arms, and head must be completely visible in order for the robot to lock onto them, but after it does, they only need to be around 50% visible. Although some occlusion of the lower half of the person is acceptable at this point, we only tested the algorithm’s abilities on short-to-medium-length grass. Its performance on landscapes with higher grass is something we plan to explore in future tests, along with other potential environmental challenges that will inevitably come up along the way. 

After confirming that the robot did well on its low speed, we moved on to test how it would do on its maximum speed of one meter per second. That is when we hit another roadblock: adjusting the speed in the code wasn’t working. Once again, Kevin and Rodolfo came through with their technical expertise and resolved the issue. Kevin offered some insight into troubleshooting the problem: “When we first tried to increase the following speed, we increased the max forward speed. We forgot that the robot is technically moving backwards when it is in person following mode, so it has a negative velocity. When we remembered this, we were able to change the ‘min_speed’ [parameter in the configuration file] to a more negative number to increase the follow speed. [It was] essentially just a sign error.” Watching the robot operate so well at its fastest was another highly encouraging moment for the team! 

To drive the point home, our Principle Investigator Professor Ricardo de Castro carried his own wagon full of wood waste alongside the robot to visually demonstrate the contrast between a physically taxing task for a human and a faster, heavy-duty robot that doesn’t experience fatigue. Check out the video below to see this in action!

Video recap of our trial runs in Catheys Valley.

We wrapped up with an obstacle avoidance test, where Kevin was the target person and Professor de Castro, standing between him and the robot, was the obstacle. Before, this set up worked perfectly on concrete when we ran it on campus, but we did not know if the new location would pose a threat to detection. To our satisfaction, the robot maneuvered around Professor de Castro to reach Kevin without any issues, proving its ability to effectively respond to unexpected obstacles in a busier environment.

With a carrying capacity of 100-150 kilos, our robot is well on its way to becoming a tool for reducing the physical burden on forestry workers. In the future, we think it is very plausible that the labor-intensive parts of forestry work like thinning, collecting debris, and hauling wood will be fully automated.

Along with celebrating this great achievement, we enjoyed talking with curious spectators about what we were testing, our follow-me technology, and the overall background and purpose of our project. Although we experienced a few hiccups, we had an overall successful trial run day that served as a productive step forward for our team!

To read more about the debugging process, check out Kevin’s guest post, “On Debugging: An Insider’s Perspective.”