Towards Terrain-Aware Safe Locomotion for Quadrupedal Robots Using Proprioceptive Sensing

This paper presents a proprioception-only framework for quadrupedal robots that combines a 2.5-D terrain estimation method with safety-critical control barrier functions to achieve robust state estimation and rigorous safety guarantees on uneven terrain.

The Gist

Imagine a four-legged robot dog trying to navigate a rugged, rocky mountain trail. Now, imagine that this robot is wearing a blindfold. It can't see the rocks, the steep drops, or the muddy puddles. It only has two ways to know where it is:

1. Proprioception: It knows where its own legs are (like you knowing your hand is raised even with your eyes closed).
2. Touch: It feels the ground when its feet hit it.

This paper is about teaching a robot to walk safely on this blindfolded, rocky terrain using only its sense of touch and body awareness, without needing expensive cameras or lasers.

Here is the breakdown of their "magic trick" using simple analogies:

1. The Problem: The "Blindfolded Hiker"

Most robots use fancy cameras (LiDAR) to "see" the ground ahead. But cameras are heavy, expensive, and useless in fog, smoke, or total darkness. If a robot relies on them, it might trip in a dusty cave or a smoky fire zone. The authors wanted to build a robot that is like a blindfolded hiker who can still walk confidently just by feeling the ground under their feet.

2. The Solution Part 1: Building a Mental Map (Terrain Estimation)

How does a blindfolded hiker know the shape of the mountain? They remember where they stepped.

• The Old Way: Previous methods were like a hiker taking a single step, guessing the slope, and then forgetting it. Or, they would take many steps and try to smooth out the memory, which often made the edges of cliffs look blurry.
• The New Way (This Paper): The robot acts like a smart cartographer.
  • Every time a foot touches the ground, the robot doesn't just record "foot here." It calculates the angle of the ground using the positions of its other legs (like drawing a triangle between three feet to guess the slope).
  • It then updates a 2.5-D map (a flat map with height data) in its brain.
  • The Analogy: Imagine the robot is painting a picture of the floor as it walks. Instead of painting jagged, messy dots, it uses a special brush that blends the dots perfectly, creating a smooth, continuous picture of the terrain. This allows it to know exactly how steep a hill is, even if it hasn't walked on that specific spot yet, by looking at the "paint" it made earlier.

3. The Solution Part 2: The "Feeling" vs. "Touch" Mix (Contact Estimation)

Sometimes, a robot's foot touches the ground, but the force is so light that its sensors think, "Oh, I'm still in the air!" This is called a "pseudo-contact." It's like stepping on a very soft pillow; you feel it, but your brain might think you're floating.

• The Fix: The robot combines its force sensors (how hard it's pushing) with its terrain map (where it thinks the ground is).
• The Analogy: It's like walking in the dark. If you feel a slight bump (force) but your memory map says "there is a wall right here," your brain says, "Okay, I'm definitely touching the wall," even if the bump was tiny. This stops the robot from thinking it's floating when it's actually stumbling.

4. The Solution Part 3: The Safety Guardian (CBF-MPC)

Now that the robot knows where the ground is, how does it not fall off a cliff?

• The Guardian Angel: The researchers added a "Safety Guardian" to the robot's brain. This guardian uses math to draw an invisible fence around the robot.
• Global Safety (The Cliff): The guardian looks at the map ahead. If it sees a steep drop-off (a cliff) coming up, it draws a "Do Not Cross" line. If the robot tries to walk toward the cliff, the guardian gently pushes the robot back, like a parent holding a child's hand at the edge of a pool.
• Local Safety (The Trip): The guardian also watches the robot's body tilt. If the ground is slanted and the robot starts to lean too far (like a person leaning on a steep hill), the guardian forces the robot to adjust its posture to stay upright, preventing it from face-planting.

5. The Results: The "Super-Blind" Robot

The team tested this on a real robot (Unitree Go1) and in simulations.

• Accuracy: By using this "touch-and-feel" map, the robot's estimate of where its body is in space became 65% more accurate. It stopped guessing and started knowing.
• Safety: When they tried to trick the robot into walking off a simulated cliff, the "Safety Guardian" kicked in. The robot stopped, turned, and walked away, saving itself from a fall.

Summary

This paper is about giving a robot super-senses. Instead of relying on expensive eyes that fail in bad weather, they taught the robot to build a perfect mental map of the world using only its feet and joints. Then, they gave it a strict "Safety Guardian" that uses this map to ensure the robot never walks off a cliff or trips over a rock.

It's the difference between a hiker who needs a flashlight to see the path, and a hiker who has memorized the path so well they can walk it in total darkness without falling.

Technical Summary

Here is a detailed technical summary of the paper "Towards Terrain-Aware Safe Locomotion for Quadrupedal Robots Using Proprioceptive Sensing."

1. Problem Statement

Quadrupedal robots face significant challenges in achieving safe locomotion over unstructured, uneven terrain, particularly in hazardous environments (e.g., rescue missions) where external sensors (LiDAR, depth cameras) may fail due to occlusion, dust, or poor lighting.

• The Core Issue: Existing methods often rely on expensive exteroceptive sensors or decouple terrain estimation from state and contact estimation. This leads to:
  • High system cost and payload.
  • Fragility in degraded perceptual conditions.
  • Lack of integration between terrain data, robot state, and contact logic, resulting in estimation errors.
  • Safety control (e.g., Model Predictive Control) that often ignores local body-terrain collision risks or relies on external sensing for global safety.
• Goal: To enable robust, safe locomotion using only proprioceptive sensors (IMUs, joint encoders, and low-cost force sensors) by creating a coupled estimation framework and a safety-critical control system.

2. Methodology

The proposed system follows a hierarchical architecture integrating a Proprioceptive Estimation Module and a CBF-MPC Control Module.

A. Proprioceptive Estimation Framework

The authors propose a coupled estimation system that simultaneously estimates terrain, contact, and robot state without post-processing.

1. Terrain Estimation (Probabilistic Fusion):
   • Generates a 2.5-D elevation map and support plane parameters ($K_1, K_2, K_3, D$) using only contact data.
   • Algorithm:
     • Plane Computation: For each leg, a triangle is formed with its two neighbors to calculate local plane parameters.
     • Map Update: A probabilistic fusion updates the global 2.5-D map. Points under the robot's support area are updated based on weighted plane confidences; points outside remain unchanged.
     • Global Plane Extraction: Principal Component Analysis (PCA) is applied to the points beneath the robot to derive a smooth, global support plane.
2. Contact Estimation:
   • Addresses "pseudo-contact" (physical contact where force is below sensor thresholds).
   • Fusion Strategy: Combines force measurements ($P_{force}$) and terrain-based position estimates ($P_{pos}$) using a weighted sigmoid function:
     $$P(c) = k_{pos}P(c)_{pos} + k_{force}P(c)_{force}$$
   • $P(c)_{pos}$ uses the estimated plane parameters to determine the likelihood of contact based on foot height relative to the terrain.
3. State Estimation:
   • Uses a Kalman Filter (KF) to estimate the Center of Mass (CoM) position and velocity.
   • Innovation: The observation vector is updated using the estimated plane parameters and contact probabilities. This corrects the leg height components in the measurement model, significantly reducing CoM estimation errors compared to decoupled methods.

B. Safety-Critical Control (CBF-MPC)

The estimated terrain and state information are fed into a Control Barrier Function (CBF)-based Model Predictive Control (MPC) framework to guarantee safety.

1. Global Safety (Hazard Avoidance):
   • Uses the 2.5-D map to identify hazardous points (steep slopes or cliffs) ahead of the robot.
   • Defines a safe set based on the distance to the nearest hazardous point projected onto the current support plane.
   • Enforces a constraint that prevents the robot from entering a zone where the slope exceeds a threshold ($\theta_{thr}$).
2. Local Safety (Body Orientation):
   • Prevents body-terrain collisions by constraining the robot's pitch and roll angles.
   • Adapts the safe set boundaries dynamically based on the estimated terrain slope ($\theta_{pitch}, \theta_{roll}$) and relaxation tolerances.
   • Ensures the robot maintains a stable orientation relative to the local ground plane.

3. Key Contributions

1. Novel Proprioceptive Terrain Estimation: A probabilistic fusion method that outputs both a smooth 2.5-D terrain map and support plane parameters without requiring external sensors or post-processing filters.
2. Coupled Estimation Architecture: A unified framework that integrates terrain, contact, and state estimation. This cross-domain coupling significantly improves estimation accuracy compared to decoupled approaches.
3. CBF-MPC Safety Framework: A control strategy that leverages proprioceptive terrain estimates to enforce both global safety (avoiding hazardous regions) and local safety (preventing body-terrain collisions) in real-time.
4. Experimental Validation: Comprehensive validation on the Unitree Go1 platform and in simulation (Gazebo), demonstrating robustness in unstructured environments.

4. Experimental Results

The authors conducted extensive simulations and real-world hardware experiments:

• Terrain Estimation:
  • Produced smooth 2.5-D maps without the "trench-like gaps" seen in point-updating baselines.
  • Achieved low angular errors for support plane estimation ($\approx 0.14^\circ$ and $0.33^\circ$ on slopes).
• Contact Estimation:
  • Successfully handled "pseudo-contact" scenarios where force sensors failed to detect contact, maintaining continuity in complex terrain.
• State Estimation:
  • 64.8% reduction in Mean Absolute Error (MAE) for base position estimation.
  • 47.2% reduction in estimation variance compared to decoupled frameworks.
• Safety Performance:
  • In simulations, the robot without CBFs entered hazardous zones and tipped over.
  • With CBFs, the robot successfully detected hazards, stopped, and navigated around them, preventing falls.

5. Significance

This work represents a major step forward in low-cost, robust legged robotics. By proving that safe, high-performance locomotion is possible using only proprioceptive sensors, the research:

• Reduces Dependency: Eliminates the need for expensive, fragile, or heavy exteroceptive sensors (LiDAR/Depth cameras).
• Enhances Reliability: Provides a solution for operation in "perceptually degraded" environments (smoke, dust, darkness) where vision-based systems fail.
• Improves Safety: Introduces a rigorous mathematical framework (CBFs) that guarantees safety at both the navigation and body-orientation levels, addressing a critical gap in current quadrupedal control literature.
• Scalability: The methodology is applicable to a wide range of quadrupedal robots, particularly those designed for search-and-rescue or exploration in harsh conditions.