Robotics engineers at ETH Zurich’s Robotics Systems Lab have upgraded a quadruped robot, allowing it to climb standard ladders quickly and easily.
Manufacturers deploying robots have long acknowledged their limitations in climbing simple ladders, which remains challenging despite significant advancements in replacing human labour. To address this gap, engineers from various companies have focused on enhancing humanoid bipedal robots, which typically climb ladders slowly and tentatively, rendering them less effective in practical scenarios.
A group of robotics engineers from the Robotics Systems Lab at ETH Zurich has taken a different approach by modifying a quadruped robot, ANYMal, enabling it to climb standard ladders swiftly and agilely.
The team observed that typical robotic “hands” or “paws” aren’t well-suited for climbing ladders. They noted that humans typically shape their hands into hooks to grip ladder rungs as they ascend. Motivated by this, the researchers developed a robotic paw capable of forming a hook-like grip, which can tighten around a rung. They employed reinforcement learning to train the robot to use its new gripping ability to climb.
To expedite the training process, the team utilised a simulation with a teacher-student method. In this setup, the “teacher” robot had access to videos showing various simulated ladder-climbing scenarios, including challenges like unstable ladders or missteps. This approach enhanced the robots’ ability to master robust climbing techniques—subsequently, multiple “student” robots learned by imitating the teacher.
After extensive training, a test robot was deployed in real-world scenarios to climb various ladders, achieving a success rate of about 90%. This robot significantly outperformed its counterparts that lacked the specialised hooked paws.
The research team aims to enhance their robots’ ladder-climbing capabilities, focusing on adapting to unstructured environments without relying on motion capture technology.
Reference: Dylan Vogel et al, Robust Ladder Climbing with a Quadrupedal Robot, arXiv (2024). DOI: 10.48550/arxiv.2409.17731
