Four-legged robot learns to sprint up ladders

Quadruped robots can sprint across gravel and climb standard stairs with ease, but standard industrial ladders have always stopped them dead in their tracks.

Now, researchers at ETH Zurich have taught their ANYmal quadruped how to scale near-vertical ladders, solving one of the nastiest navigation headaches in industrial robotics.

Hooks and reinforcement learning

Historically, teaching robots to climb ladders has been an agonizing exercise. Most attempts relied on bipedal humanoids on custom-built ladders, moving at an excruciatingly slow pace that made real-world deployment useless.

ETH Zurich took ANYmal, the quadruped developed by its spinoff ANYbotics, and tackled the physical and algorithmic constraints head-on.

First, they outfitted the four-legged robot with specialized hook-like end effectors designed to latch securely onto standard ladder rungs.

The real secret sauce, though, is reinforcement learning. Instead of hand-coding every foot placement, the neural control policy trains the robot to feel its way up and dynamically adjust to different ladder geometries and spacing.

If the robot slips, misjudges a rung, or mis-times a step, the controller catches the error in real time and repositions the limb without stalling the entire ascent.

The results speak for themselves. The system achieved a 90% success rate on steep ladder angles ranging from 70 to 90 degrees.

Even more striking is the pace: ETH Zurich reports that ANYmal climbed ladders 232 times faster than current state-of-the-art robotic systems.

Why it matters

For industrial operators in chemical plants, offshore rigs, and manufacturing hubs, vertical access has been the ultimate barrier for autonomous inspection.

Deploying drones indoors poses safety risks, and sending human inspectors up rickety catwalks and vertical ladders remains dangerous. If quadrupeds can handle standard industrial ladders without requiring infrastructure redesigns, autonomous inspection coverage jumps dramatically.

We think this approach highlights a smart philosophical shift in robotics. While the broader industry remains fixated on shoehorning complex bipedal humanoids into every industrial task, ETH Zurich proved that a purpose-built quadruped with clever end effectors and reinforcement learning can run circles around humanoids.

Still, a 90% success rate in controlled settings leaves a 10% failure window. Falling from a vertical steel ladder onto active machinery or concrete means a totaled robot and severe safety hazards. Before this rolls out to live refineries, ANYbotics will need bulletproof physical fail-safes and testing on greasy, corroded rungs.