💻 computer science

Global-Local Attention Decomposition for Terrain Encoding in Humanoid Perceptive Locomotion

This paper introduces Global-Local Attention Decomposition (GLAD), a novel terrain encoding framework that separates broad context awareness from precise foothold selection to enable robust, zero-shot sim-to-real perceptive locomotion for humanoid robots on sparse and constrained terrains.

Original authors: Shengcheng Fu, Yang Zhang, Zhanxiang Cao, Liyun Yan, Yizhi Chen, Yunpeng Yin, Yue Gao

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shengcheng Fu, Yang Zhang, Zhanxiang Cao, Liyun Yan, Yizhi Chen, Yunpeng Yin, Yue Gao

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 room filled with scattered stepping stones, narrow paths, and sudden gaps. To do this successfully, the robot needs to solve two very different problems at the same time:

The Big Picture: It needs to look far ahead to see the general layout of the room (e.g., "Is there a path forward, or is it a dead end?").
The Fine Print: It needs to look very closely at the specific spot where its foot is about to land to ensure that exact stone is solid and reachable.

For a long time, robot "brains" (neural networks) tried to do both of these things with a single, jumbled focus. It was like trying to read a map while simultaneously trying to thread a needle; the brain would get confused, diluting the important details.

This paper introduces a new method called GLAD (Global-Local Attention Decomposition) to fix this. Here is how it works, using simple analogies:

The Problem: The "Overloaded Chef"

Think of a traditional robot encoder as a chef who is trying to cook a complex meal while also reading a novel. They are trying to process everything at once.

They look at the whole kitchen (the global view).
They look at the specific ingredient in their hand (the local view).
The Result: They get distracted. They might miss that the ingredient is slightly rotten because they were too focused on the story, or they might miss the story because they were too focused on the ingredient. In robot terms, this leads to clumsy walking or falling off stepping stones.

The Solution: GLAD (The "Specialized Team")

The authors propose splitting the robot's brain into two specialized assistants who work together but have distinct jobs.

1. The "Scout" (Global Attention Branch)

Job: This assistant stands on a hill and looks at the entire landscape.
How it works: It uses a technique called "attention pooling" to take a quick, broad summary of the terrain ahead. It doesn't worry about the texture of every single rock; it just wants to know, "Is there a path? Are there big gaps? Is the ground generally safe?"
Analogy: It's like a tour guide giving you a general overview of the city before you start walking.

2. The "Spotter" (Local Attention Branch)

Job: This assistant is the robot's eyes, zooming in on the specific spot where the foot is about to land.
How it works: This is the clever part. Instead of looking at every rock in the path, the Spotter uses the robot's current body position (is it leaning left? moving fast?) to filter out the rocks that don't matter. It throws away 80% of the visual data and only keeps the top few rocks that are actually relevant to the next step. It then analyzes those few rocks in extreme detail.
Analogy: It's like a sniper focusing only on the target, ignoring the background noise.

Why This Matters

By separating these two jobs, the robot doesn't get confused.

The Scout ensures the robot doesn't walk off a cliff or into a wall.
The Spotter ensures the robot places its foot precisely on a small, wobbly stone.

Because the Spotter ignores irrelevant data, the robot's brain works faster and learns to walk better. It doesn't have to waste energy processing the whole world; it just processes what it needs for the next step.

The Results: Walking the Walk

The researchers tested this on a real robot (a Unitree G1) and in computer simulations.

The Test: They threw the robot into difficult scenarios: narrow stepping stones, wide gaps (up to 70cm), stairs with missing steps, and paths cluttered with obstacles.
The Outcome: The robot using GLAD could walk across these terrains smoothly. It could even follow narrow paths and dodge obstacles just by being told "walk forward," without needing a separate computer to plan the route.
Real-World Success: The robot learned in a simulation and then walked perfectly in the real world without any extra tuning. It used only its onboard laser scanner (LiDAR) to "see" the ground.

Summary

In short, this paper teaches a robot to stop trying to do everything at once. Instead, it gives the robot a Scout to see the big picture and a Spotter to focus on the immediate step. This simple division of labor allows the robot to walk confidently over tricky, uneven ground where previous robots would have stumbled.

Technical Summary: Global-Local Attention Decomposition for Terrain Encoding in Humanoid Perceptive Locomotion

Problem Statement
While reinforcement learning (RL) has advanced humanoid locomotion, perceptive policies struggle in sparse-foothold and constrained environments. Success in these scenarios requires two distinct perceptual capabilities: broad awareness of the surrounding terrain context and precise selection of feasible footholds. Conventional terrain encoders often entangle these objectives within a single mechanism, leading to the dilution of fine-grained spatial cues, weakened spatial correspondence between features and reachable regions, and increased training overhead. Existing attention-based methods (e.g., AME, AME-2) attempt to fuse proprioception with exteroception but typically apply a single attention mechanism to the full feature set, failing to explicitly separate global context aggregation from state-conditioned local encoding.

Methodology: Global-Local Attention Decomposition (GLAD)
The authors propose GLAD, a coarse-to-fine terrain encoder designed to explicitly decouple these perceptual roles. The architecture processes a robot-centric elevation map through the following stages:

Spatially Aligned Local Feature Extraction: A two-layer Convolutional Neural Network (CNN) extracts local features from the 3D elevation map. Unlike prior works that convolve only height values, GLAD convolves the full 3D coordinates to embed positional cues directly into the features. The first layer performs strided downsampling to decouple sensing resolution from feature resolution, reducing sequence length while maintaining spatial alignment with the terrain.
Global Attention Branch: This branch utilizes attention pooling to aggregate all local features into a compact "surrounding terrain context" vector. This vector summarizes the broad traversability and geometric layout ahead of the robot without focusing on specific footholds.
Local Attention Branch: This branch performs state-conditioned sparsification and encoding. It fuses the global context vector with a proprioceptive embedding to generate a state-conditioned query. This query scores local features, retaining only the top- $K$ candidates (sparsification) relevant to the robot's current motion state. A Multi-Head Attention (MHA) mechanism then processes these sparse features to produce a fine-grained, foothold-relevant geometric encoding.
Unified Representation: The global context vector and the local foothold-relevant feature are concatenated to form the final terrain representation fed into the locomotion policy.

Key Contributions

Explicit Decomposition: The paper introduces GLAD, which separates attention into global terrain-context aggregation and state-conditioned local foothold encoding. This allows the policy to jointly reason about broad traversability and precise support-region geometry without architectural entanglement.
Efficiency and Interpretability: By combining decomposition with state-conditioned local feature sparsification, the method reduces the computational cost of applying dense attention to the full terrain feature set. The resulting attention behaviors are complementary and interpretable: global attention focuses on forward traversability, while local attention highlights specific foothold regions.
Robust Zero-Shot Transfer: The method is validated on a Unitree G1 humanoid robot using only onboard LiDAR-based elevation mapping. It achieves robust zero-shot sim-to-real transfer across diverse sparse-foothold and obstacle-rich domains without pre-mapping or external motion capture.

Experimental Results

Simulation: GLAD was tested against baselines (AME, AME with global context) and ablation variants on four unseen test terrains, including composite stepping stones, random pillars, and narrow sinusoidal paths. GLAD achieved the highest success rates (e.g., 95.90% on narrow paths vs. 18.95% for GLAD without local attention).
Attention Analysis: Quantitative metrics showed GLAD achieves the smallest attention-weighted distance to actual footholds and the highest attention concentration ( $A_{peak}$ ). Notably, GLAD's local attention concentration adapts to terrain difficulty (increasing from 0.257 on continuous terrain to 0.447 on discrete terrain), whereas baselines showed less adaptability.
Training Efficiency: The coarse-to-fine design reduced training time to 1.00 days on a single GPU, compared to 1.36 days for AME and 1.50 days for AME with global context.
Real-World Deployment: On the Unitree G1, the policy successfully traversed alternating stepping stones, staircase-gap composites, wide gaps (up to 70 cm), and dense obstacle fields. The robot exhibited emergent behaviors, such as autonomously following narrow paths and avoiding obstacles under simple velocity commands, without hierarchical planners.

Significance and Claims
The paper claims that GLAD addresses a fundamental limitation in perceptive locomotion: the entanglement of global context and local precision in existing encoders. By explicitly separating these roles, the method enables more robust, terrain-responsive legged locomotion. The authors emphasize that the learned policy exhibits "emergent terrain-responsive behaviors," allowing the robot to navigate complex, constrained environments using only simple velocity commands and onboard sensing. The work demonstrates that high-quality terrain representation, rather than complex reward shaping or handcrafted foothold rewards, is sufficient to support reliable locomotion on challenging sparse-foothold terrain.

Limitations
The authors acknowledge limitations regarding the sensing range of the local elevation map, reconstruction noise, and the computational overhead of real-time elevation mapping. Future work is suggested to investigate more efficient mapping pipelines, temporally extended perception, and multi-scale terrain representations.