Amazon's LadderMan Teaches Humanoid Robots to Climb at Near-Human Speed
Amazon FAR's LadderMan teaches a humanoid robot to climb diverse real-world ladders and perform manipulation tasks using a single motion capture reference and zero-shot sim-to-real transfer.
- LadderMan is a system from Amazon FAR, Berkeley, and Stanford that enables a Unitree G1 humanoid to climb diverse real-world ladders and perform manipulation tasks with no hardware modifications.
- A two-stage pipeline trains expert policies from a single motion capture clip, then distills them into a unified depth-based visuomotor policy via hybrid DAgger + RL.
- A vision foundation model (Fast-FoundationStereo) replaces manual depth randomization, producing clean real-world depth from a RealSense camera for zero-shot sim-to-real transfer.
- LadderMan achieves over 95% success across a range of ladder geometries in simulation and climbs at ~3.4 s/rung, comparable to a human's ~3.2 s/rung.
- A dual-agent manipulation policy decouples lower-body balance from upper-body teleoperation, enabling tasks like adjusting paintings and replacing light bulbs while on the ladder.
- Code and models are open-sourced at github.com/amazon-far/LadderMan; current limits include no support for vertical ladders and no dexterous end-effectors.
Getting a humanoid robot to walk on flat ground is hard enough. Getting it to climb a ladder, perceive the rungs in real time, and then adjust a wall painting while perched at the top? That is a different problem entirely. Researchers from Amazon FAR, UC Berkeley, Stanford, USC, and CMU have released LadderMan, a system that does exactly that on a Unitree G1 humanoid, with no hardware modifications and no fine-tuning on the target ladder.
Why ladders are so brutally hard
Ladder climbing sits at the intersection of everything that makes humanoid control difficult. Unlike walking on flat ground or stairs, ladder climbing requires precise and sustained coordination between upper and lower body contacts under sparse, discontinuous, and narrow footholds and handholds. Miss a rung by a centimeter and the robot falls. The problem is compounded by the real world: ladders vary in inclination angle, rung spacing, rung width, and material. A policy trained on one ladder will almost certainly fail on another.
Prior work on ladder climbing largely relies on carefully engineered model-based control without incorporating perception into the control loop, and requires accurate environment modeling, task-specific tuning, and specialized hardware design. These approaches typically exhibit slow climbing speeds, limited robustness, and poor generalization beyond specific ladder configurations. Learning-based approaches existed for quadruped robots, but humanoids introduce a harder version of the problem: heterogeneous limbs, a much taller and less stable body, and the need to coordinate hands and feet simultaneously.
A four-stage pipeline from one motion clip
LadderMan's core insight is to decompose the problem into stages rather than trying to learn everything end-to-end. The climbing policy is built on a scalable two-stage learning pipeline, where hybrid motion tracking is used to learn multiple climbing experts from a single reference motion, and these experts are distilled into a unified depth-based visuomotor climbing policy via hybrid imitation and reinforcement learning.
Here is how the four stages work in practice:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
