Robust Ladder Climbing with a Quadrupedal Robot

Authors: Dylan Vogel, Robert Baines, Joseph Church, Julian Lotzer, Karl Werner, Marco Hutter · IROS 2025

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

Hooked end-effectors and a teacher-student RL policy let an ANYmal D climb ladders inclined 70°–90° zero-shot, with 90% hardware success and an average speed of 0.44 m/s.

Why it matters

Ladders are common in industrial infrastructure and ladder falls are a major occupational hazard (the paper cites an estimated 500,000 ladder-related injuries per year in the US), so a quadruped that can climb them extends inspection to places it could not reach before. The work also shows that a modest change of foot shape, trained together with the policy, separates fragile from reliable climbing under noise and disturbances.

Problem

Quadrupeds are used for industrial inspection but cannot reliably climb ladders, which keeps them out of dangerous locations and leaves that work to people. Their ball-shaped or flat feet cannot anchor on rungs, climbing needs whole-body coordination at steep inclines, and ladders vary in length, width, rung spacing, rung shape and incline. Earlier robotic ladder climbing, mostly by humanoids, handled only vertical ladders without perturbations and was very slow (26 mm/h for one humanoid, up to two minutes per rung for quadrupeds).

Proposed solution

The authors fit an ANYmal D with four hooked end-effectors (aluminium cores, 3D-printed shells, concave surfaces that settle on round rungs and protrusions for pushing and pulling) and train a position-command tracking policy in LeggedGym. A teacher policy trained with Interior-point Policy Optimization uses privileged state and a height scan; a student with a recurrent belief encoder is distilled to use only noisy proprioception, IMU history and an estimated ladder state and pose. The student outputs joint position targets at 50 Hz, tracked by a 400 Hz PD controller, and is deployed zero-shot.

Key insights

Foot shape matters under uncertainty: without noise, ball and hook feet both exceed 99% success in simulation, but with noise and pushes hooks reach 96% versus 81% for ball feet.
The hooks let the robot hang in tension with its centre of mass outside the support triangle, switching from pushing down on a rung to pulling when it is disturbed.
Raw IMU history replaced the ANYmal state estimator's base velocity, which showed linear errors of up to 1 m/s during climbing.
Dropping sampled command times in favour of a dense tracking reward and episodes of up to 10 s improved success and produced emergent retry behaviour after slips.
One policy handles both walking and climbing, and its emergent gait resembles human ladder climbing: hind legs give upward thrust while forelegs mainly stabilise.

Main results

Simulation over 50 ladder configurations with 3,072 agents, noise and random pushes: 96% average success with hooks versus 81% with ball feet.
Hardware, zero-shot, on a 1.8 m ladder with five cylindrical rungs: 4 of 4 climbs at 70°, 4 of 4 at 80° and 2 of 3 at 90°, reported as 90% overall.
Average climbing speed 0.44 m/s (0.51 body lengths/s, standard deviation 0.16 m/s) over ten successful configurations, 232× faster than the previous state-of-the-art quadruped.
Sim-to-real agreement at 80°: base position and orientation RMSE of 11 cm and 0.18 rad over the whole trajectory.
Climbing continued while the robot was pulled by a rope tied to the front or back of the chassis or to the front left foot.
Training took about 4.5 days on one RTX 3090 (teacher 15,000 epochs, student 5,000, 4,096 parallel environments); hook collision bodies made training about 30% slower.

Limitations

The third 90° trial failed because unmodelled standoffs on the robot's base hit the top rung; the authors stopped 90° tests because of this sim-to-real gap.
Ladder pose and incline come from a motion-capture system and other ladder parameters are measured by hand; replacing motion capture with sensing such as depth cameras is future work.
Only ascent is demonstrated; climbing down is stated as future work.
The hook design was chosen heuristically and is not optimised.
In simulation, rungs of 3.5 cm radius, larger than the hook's 2.75 cm opening, caused a steady drop in success at steeper inclines.

Field impact

The paper was an IROS 2025 award finalist (program session WeAT1) and has 25 citations. Neither the paper nor its project website states a release of code or hook design files.

Abstract

Quadruped robots are proliferating in industrial environments where they carry sensor suites and serve as autonomous inspection platforms. Despite the advantages of legged robots over their wheeled counterparts on rough and uneven terrain, they are still yet to be able to reliably negotiate ubiquitous features of industrial infrastructure: ladders. Inability to traverse ladders prevents quadrupeds from inspecting dangerous locations, puts humans in harm's way, and reduces industrial site productivity. In this paper, we learn quadrupedal ladder climbing via a reinforcement learning-based control policy and a complementary hooked end-effector. We evaluate the robustness in simulation across different ladder inclinations, rung geometries, and inter-rung spacings. On hardware, we demonstrate zero-shot transfer with an overall 90% success rate at ladder angles ranging from 70° to 90°, consistent climbing performance during unmodeled perturbations, and climbing speeds 232× faster than the state of the art. This work expands the scope of industrial quadruped robot applications beyond inspection on nominal terrains to challenging infrastructural features in the environment, highlighting synergies between robot morphology and control policy when performing complex skills. More information can be found at the project website: https://sites.google.com/leggedrobotics.com/climbingladders.

Open Source Robotics