ETH Zurich’s ANYMal robot aces ladders with its custom-built hooked feet

Engineers at ETH Zurich’s Robotics Systems Lab modified the ANYbotics ANYMal quadruped robot to efficiently climb standard ladders.

The upgrade enhances its mobility in complex environments, making it more versatile for tasks like inspection and rescue operations.

The team developed custom hook-like paws and used reinforcement learning with simulation to teach ANYMal to climb ladders, enhancing its climbing skills through a teacher-student training approach.

After training, the robot was tested in real-world ladder climbs, achieving 90 percent success. It significantly outperformed unmodified versions of the same robot, proving the effectiveness of the hooked feet design.

Advanced ladder climbing

Quadruple robots are becoming more commonplace in industrial settings. They can carry sensor suites and act as independent platforms for inspection.

Even while legged robots are superior to wheeled counterparts in difficult and uneven environments, they are still unable to consistently navigate one of the most common elements of industrial infrastructure: ladders.

The incapacity to climb ladders hinders quadrupeds from examining hazardous areas, endangers humans, and lowers production at industrial sites.

Previous research on robotic ladder climbing primarily focused on humanoid robots, which showed slow climbing speeds and limited adaptability to diverse ladder configurations.

Studies on quadrupedal robots also faced similar constraints, demonstrating slow ascent on only vertical ladders. Beyond ladder climbing, robust locomotion in quadrupeds has been achieved in challenging terrains using model-based methods, but these are sensitive to disturbances.

Reinforcement learning (RL), on the other hand, has shown real-world robustness but hasn’t yet addressed ladder climbing.

With the new research, ETH engineers present a new RL framework for robust climbing, a custom hook design for secure gripping, and extensive testing, achieving the fastest and most adaptable ladder climbing for quadrupeds.

Hooked for success

Past experiments led the team to conclude that robotic hands or paws are not ideal for ladder climbing, as humans naturally form their hands into hooks to grip each rung.

They created a special paw with a hook-like mechanism that could tightly grasp and clasp onto ladder rungs in order to address this. The robot was then trained to use these hooks for climbing via reinforcement learning .

They used a privileged teacher-student paradigm to mimic ladder-climbing events in order to expedite the training. Equipped with complete sensory data, the instructor robot acquired the ability to climb in a variety of settings, overcoming obstacles such as shaky ladders and misplaced stairs.

Then, by mimicking the teacher, a number of student robots were educated to have strong ladder-climbing abilities. To improve flexibility, extensive simulations were conducted with difficult terrain and randomly generated surroundings. This approach enabled the robots to develop highly effective climbing skills more quickly and efficiently.

“On hardware, we demonstrate zero-shot transfer with an overall 90 percent success rate at ladder angles ranging from 70° to 90°, consistent climbing performance during unmodeled perturbations, and climbing speeds 232x faster than the state-of the-art,” said the team, in the study abstract.

Even with interruptions, a detailed assessment in simulation revealed a 96 percent success rate for ladder ascending. Because of the stability the hook end-effector offered, the robot was able to climb steeper ladders and deal with unforeseen obstacles.

Future research will incorporate other sensors, such as depth cameras, to enable quadrupeds to climb both up and down ladders . The team claims that the existing hook design is not optimal, creating new research opportunities to enhance the shape and control of the robot in tandem for enhanced performance in a variety of activities.

The details of the team’s research were published in the arXiv preprint server.