🤖 AI

CMR: Contractive Mapping Embeddings for Robust Humanoid Locomotion on Unstructured Terrains

This paper introduces the Contractive Mapping for Robustness (CMR) framework, which enhances humanoid locomotion on unstructured terrains by mapping noisy observations into a contractive latent space via contrastive learning and Lipschitz regularization, thereby theoretically bounding return gaps and empirically outperforming existing algorithms under significant sensor noise.

Original authors: Qixin Zeng, Hongyin Zhang, Shangke Lyu, Junxi Jin, Donglin Wang, Chao Huang

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qixin Zeng, Hongyin Zhang, Shangke Lyu, Junxi Jin, Donglin Wang, Chao Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a humanoid robot trying to walk through a chaotic, messy room filled with uneven floorboards, random obstacles, and flickering lights. To the robot, the world is a blur of confusing data: its sensors are noisy, its "eyes" (cameras and height maps) are glitching, and its internal sense of balance is fuzzy. In this chaos, a normal robot might trip, stumble, or freeze because it overreacts to every tiny glitch in its data.

This paper introduces a new method called CMR (Contractive Mapping for Robustness) to help these robots walk steadily even when their senses are failing. Here is how it works, explained through simple analogies:

1. The Problem: The "Echo Chamber" of Noise

Imagine you are trying to have a conversation in a room where every time someone speaks, a loud echo repeats their words, but slightly distorted. If you try to react to the echo, the distortion gets worse and worse with every repetition. Eventually, you can't understand anything.

In robotics, this is what happens with "observation noise." If a robot's sensors are slightly off (like a height map that says a step is 1 inch high when it's actually 2 inches), a standard robot might overcorrect. That overcorrection causes a new error, which causes another overcorrection, leading to a fall. The paper calls this an "amplification" of errors.

2. The Solution: The "Sponge" Effect

The authors propose a new way for the robot to process information. Instead of reacting directly to the noisy, glitchy data, the robot first passes that data through a special "filter" or "sponge" called a Contractive Embedding.

Think of this sponge like a shock absorber in a car.

Without the sponge: If you hit a bump, the car bounces wildly.
With the sponge: If you hit a bump, the shock absorber soaks up the energy. The car might wobble a little, but the wobble gets smaller and smaller until it disappears.

CMR does exactly this for data. It takes the messy, high-dimensional sensor data (like a 15x15 grid of height information) and squeezes it into a simpler, "latent" space. In this new space, small errors (noise) naturally shrink over time instead of growing. If the robot makes a tiny mistake in its internal map, the math of CMR ensures that mistake gets "squashed" rather than amplified as the robot takes the next step.

3. How It Learns: The "Tightrope" and the "Map"

To build this special sponge, the researchers used two clever training tricks combined into one system:

The "Lipschitz" Rule (The Tightrope): Imagine a tightrope walker who promises that if they move their foot one inch, their body will never move more than one inch. They won't suddenly fling themselves three feet to the left. This is called a "Lipschitz constraint." The robot is trained to promise that small changes in what it sees will only lead to small, controlled changes in how it moves. This prevents wild, jerky reactions to sensor glitches.
Contrastive Learning (The Map): The robot also needs to know what is important. Imagine you are in a foggy forest. You need to know that a tree is a tree, even if the fog makes it look blurry. The robot uses "contrastive learning" to make sure that two similar views of the ground (even if one is noisy) are still recognized as the same type of terrain. It keeps the "shape" of the world intact while smoothing out the noise.

4. The Results: Walking on a Tightrope in a Storm

The authors tested this on a Unitree G1 humanoid robot in a computer simulation. They threw all kinds of chaos at it:

Noisy Sensors: They added random static to the robot's vision and balance sensors.
Hard Terrains: They made the robot walk on uneven ground, stepping stones, and balance beams.
Random Commands: They told the robot to walk at random speeds and directions.

The Outcome:

Other robots (using standard methods) tended to stumble, waste a lot of energy flailing their limbs, or stop moving entirely when the noise got high.
The CMR robot kept walking. It traveled further, used less energy, and moved much more smoothly. Even when the noise was extreme, the robot didn't panic; it just kept moving forward.

5. Why This Matters (According to the Paper)

The paper claims this is the first time the mathematical concept of "Contraction Mapping" (the idea that things get closer together over time) has been successfully applied to teaching robots how to walk.

They proved mathematically that if a robot uses this method, the gap between how well it should perform and how well it actually performs is strictly limited, even when the sensors are terrible. It's like having a guarantee that no matter how bad the weather gets, your car's suspension will keep you from hitting the roof.

In short: CMR teaches a robot to ignore the "static" in its sensors and focus on the steady rhythm of walking, ensuring it doesn't trip over its own confusion.

Technical Summary: Contractive Mapping Embeddings for Robust Humanoid Locomotion on Unstructured Terrains (CMR)

Problem Statement

Robust disturbance rejection remains a critical challenge in humanoid locomotion, particularly within unstructured environments characterized by terrain irregularities, actuation perturbations, and unreliable sensing. While perception data (e.g., height maps) enhances terrain awareness, sensor noise and sim-to-real gaps often destabilize control policies. Existing approaches face significant limitations:

Model-based controllers (e.g., ZMP, LIP) rely on restrictive assumptions (fixed CoM height, simplified dynamics) and struggle with measurement uncertainty.
Learning-based methods often fail to maintain robustness when exteroceptive sensors introduce noise, especially on complex terrains like stairs or balance beams.
Current robustness techniques, such as low-pass filters or smoothness rewards, require extensive tuning, reduce exploration capacity, and often fail to generalize across diverse robotic systems.
Existing sensitivity control methods (e.g., Lipschitz constraints via gradient penalties) are often limited to basic walking behaviors on low-complexity terrains.

The core problem addressed is the lack of a systematic method to ensure robustness against observation noise in complex, high-degree-of-freedom (DOF) humanoid locomotion tasks without sacrificing task performance or requiring prohibitive engineering overhead.

Methodology

The authors propose Contractive Mapping for Robustness (CMR), a framework that maps high-dimensional, disturbance-prone observations into a latent space where local perturbations are attenuated over time.

Theoretical Foundation

The approach is grounded in the Contraction Mapping Theorem. The authors derive theoretical bounds on the "return gap" (performance degradation) under observation noise:

Standard Systems: In systems with Lipschitz constants $L_f \geq 1$ , observation noise can lead to exponential error growth over the time horizon $H$ , causing severe performance degradation.
Contractive Systems: By mapping states into a latent space with a contraction factor $0 < \kappa < 1$ , the framework ensures that differences between trajectories shrink over time. Theoretical analysis (Theorem 2) establishes that the return gap is bounded by $O(\frac{\eta}{1-\kappa})$ , where $\eta$ bounds the expected action discrepancy. Crucially, this bound is independent of the horizon length $H$ , offering superior long-horizon robustness compared to non-contractive approaches.

Framework Architecture

The CMR framework integrates three key components into a standard Deep Reinforcement Learning (RL) pipeline (specifically Proximal Policy Optimization, PPO):

Contrastive Representation Learning: Utilizes an InfoNCE loss to preserve semantic discrimination, ensuring the latent space retains task-relevant geometry and prevents excessive information loss.
Lipschitz Regularization: Introduces a specific loss term ( $L_{Lipschitz}$ ) to explicitly constrain the evolution of differences in the embedding space, enforcing the contraction property ( $\kappa < 1$ ).
Joint Optimization: The total training objective combines the PPO loss, the contrastive loss, and the Lipschitz regularization:
$L_{CMR} = L_{InfoNCE} + \lambda \cdot L_{Lipschitz} + L_{PPO}$
where $\lambda$ balances semantic discrimination and dynamic contraction.

Implementation Details

Observations: The system processes noisy proprioceptive data (joint positions/velocities, base pose) and exteroceptive data (egocentric elevation maps).
Noise Injection: During training, noise is injected into perception observations to simulate real-world degradation.
Compatibility: The formulation acts as an auxiliary loss term, requiring minimal modification to existing deep RL pipelines.

Key Contributions

Novel Framework: Introduction of CMR, a contractive-embedding framework that systematically attenuates observation perturbations while preserving task-relevant structure.
Theoretical Advancement: The first application of the contraction mapping theorem to learning-based humanoid locomotion, providing rigorous bounds on noise-induced return gaps.
Practical Integration: Development of a simple, effective training system that seamlessly integrates with modern deep RL pipelines, enabling robust control with minimal technical effort.

Experimental Results

Extensive experiments were conducted on the Unitree G1 humanoid robot in simulation (Isaac Gym and MuJoCo) across diverse terrains (uneven ground, stepping stones, balance beams) and noise levels.

Noise Rejection: Under high noise conditions (Noise Case I and III-V), CMR significantly outperformed baselines (HIM, LCP, Naive PPO). It maintained stable locomotion, achieved longer travel distances, and exhibited lower joint power consumption and smoother control actions.
Generalization (Sim-to-Sim): In zero-shot transfer scenarios from Isaac Gym to MuJoCo under noisy disturbances, CMR demonstrated superior command tracking and energy efficiency compared to the HIM baseline.
Ablation Studies: Analysis of hyperparameters ( $\tau$ , $\lambda$ , $\kappa$ ) confirmed a trade-off between trajectory accuracy and energy consumption, validating the chosen default configuration as a balanced compromise.
Latent Space Visualization: t-SNE visualizations revealed that CMR's latent trajectories exhibit significantly less deviation under noise compared to baselines, confirming the self-stabilizing properties of the contractive space.

Significance and Claims

The paper claims that CMR represents a significant step forward in robust humanoid control by:

Systematically addressing noise robustness in complex locomotion tasks, an area previously identified as an open research problem.
Providing theoretical guarantees that limit performance degradation due to observation noise, moving beyond heuristic tuning.
Enabling robust sim-to-real transfer potential by learning contractive state embeddings that provably limit noise propagation.

The authors note that while the current work focuses on simulation, they are verifying transfer performance through real-robot experiments to strengthen practical credibility. Future work aims to incorporate multi-modal perception (e.g., depth cameras) to handle even more precise and complex terrains.