The group has written a paper describing their efforts and results and has posted it on the arXiv preprint server.
For many years, manufacturers who have deployed robots in their facilities have noted that while robots have come a long way in replacing human laborers, one area where they fall short is climbing simple ladders. In response, engineers from several companies have attempted to give robots such an ability, almost all of which have been humanoid bipedal type robots.
Such robots are typically very slow and tentative and not very useful in a real-world environment. In this new effort, the research team has modified a standard working quadruped robot known as ANYMal in a way that allows it to very quickly and nimbly climb standard ladders.
The team noted that most robot "hands" or "paws" are not very conducive to ladder climbing. Humans, they note, form their hands into a hook and then grip each rung as they proceed upward. Therefore, the research team started by designing a custom paw with a hook-like gripping ability—one that allowed for clenching once a hook was made. They then used reinforcement learning to teach the robot how to use its hooks to climb a ladder.
To speed up the training, the researchers used a simulation where a privileged teacher–student approach was employed, where the teacher was given access to observational videos of simulated robots climbing ladders in a variety of environments and where problems were encountered, such as a jiggling ladder or a missed step. Such training allowed for learning robust climbing skills. Multiple student robots were then taught by allowing them to mimic the teacher.
Once they had a test robot thoroughly trained, the research team set it loose in a real-world environment where it was asked to climb a variety of ladders. The researchers found it to be successful approximately 90% of the time. They also found their modified robot greatly outperformed the same type of robot without the hooked feet.
The research team plans to continue their work with ladder climbing, hoping to add more features such as the ability to traverse ladders in unstructured environments without the need for motion capture equipment.
