💻 computer science

Energy Prediction on Sloping Ground for Quadruped Robots

This paper presents a simple, sensor-based energy model for quadruped robots that predicts locomotion costs on sloping terrain by analyzing the influence of slope angle and heading orientation, enabling effective path-level energy planning in unexplored outdoor environments.

Original authors: Mohamed Ounally, Cyrille Pierre, Johann Laconte

Published 2026-03-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohamed Ounally, Cyrille Pierre, Johann Laconte

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 you are a hiker trying to cross a mountain. You have two choices:

The Direct Route: Walk straight up the steepest part of the hill. It's short, but your legs will burn out fast.
The Zig-Zag Route: Walk a longer path, cutting across the hill at an angle. It takes more steps, but it feels much easier on your energy.

For a wheeled robot (like a car), this is easy to calculate: just look at the distance and the steepness. But for a quadruped robot (a robot with four legs, like a dog), the math is much harder. Their legs have to lift, swing, and land repeatedly. Every time a foot hits the ground, it's a tiny collision that wastes energy.

This paper is about teaching these legged robots to be "smart hikers" that know exactly how much battery power they will burn based on where they are going and how they are facing the hill.

Here is the breakdown of their discovery, using some everyday analogies:

1. The Problem: The "Black Box" of Legged Energy

Imagine you are driving a car. You know that going uphill uses more gas, and going downhill uses less. It's a simple relationship.

Now, imagine a robot dog. If it walks straight up a hill, it works hard. If it walks sideways across the hill, it has to fight gravity and keep from slipping, which is exhausting. If it walks down, it doesn't just "coast" like a car; it has to actively control its legs to not fall over, which still uses energy.

The researchers realized that for these robots, direction matters just as much as distance. A long, gentle path might actually save more battery than a short, steep one. But until now, robots didn't have a simple way to predict this.

2. The Solution: A "Smart Battery Map"

The team created a simple "energy map" for robots. Think of it like a weather forecast, but instead of rain, it predicts battery drain.

No Fancy Tools Needed: Usually, to measure how much energy a robot uses, you need expensive sensors on every joint to measure muscle tension. The researchers found a clever shortcut. They just looked at the robot's battery (how fast the voltage drops) and its speed. It's like estimating how hard you are running just by looking at how fast your heart rate goes up, without needing a lab test.
The "Force" Formula: They figured out that the energy cost is basically a combination of two things:
1. The Slope: How steep is the hill?
2. The Heading: Is the robot facing straight up, straight down, or sideways?

3. The Big Discoveries (The "Aha!" Moments)

Through experiments with a real robot dog (a Unitree B1) on grassy hills and ramps, they found three surprising things:

The "Sideways Struggle": Walking straight up a hill is hard. Walking straight down is easier. But walking sideways across a hill is the most exhausting of all! It's like trying to walk sideways on a slippery staircase; you have to fight to stay balanced, which burns extra energy.
Downhill Isn't Free: You might think walking down a hill gives you a "free ride." For a robot, it's not free. It has to brake constantly with its legs to avoid tumbling. It's like walking down a steep hill in high heels; you aren't climbing, but you still have to work hard to stay upright.
Energy Adds Up: If you walk a path made of three different segments, the total energy used is just the sum of the three parts. This is great news for robot planners because it means they can calculate the cost of a long journey by just adding up the costs of small steps.

4. Why This Matters

Imagine a robot sent to a farm to check crops, or a rescue robot sent into a disaster zone.

Without this model: The robot might pick the shortest path, run out of battery halfway, and fail the mission.
With this model: The robot can say, "That direct path up the hill will kill my battery. Instead, I'll take this slightly longer, zig-zag path across the slope. It uses less energy, and I'll make it to the finish line."

The Bottom Line

This paper gives robots a "sixth sense" for energy. It allows them to stop guessing and start planning. By understanding that facing the right way on a slope is just as important as the slope itself, these robots can work longer, do more tasks, and be more efficient in the real world.

It's the difference between a hiker who just runs up the mountain and collapses, and a hiker who knows exactly how to pace themselves to reach the summit with energy to spare.

1. Problem Statement

Legged robots are increasingly deployed in demanding outdoor environments (agriculture, mining, disaster response) where energy endurance is a critical constraint. While wheeled vehicle energy models are well-established (often relating cost to distance and slope), predicting energy consumption for quadruped robots is significantly more complex due to:

Multi-body dynamics: The interaction of four legs, coordinated gaits, and repeated ground contacts.
Heading dependency: Unlike wheeled vehicles where slope magnitude is the primary factor, legged robots exhibit different energy costs depending on their orientation (heading) relative to the slope (e.g., climbing straight up vs. traversing laterally).
Sensor limitations: Existing high-fidelity models often require specialized instrumentation (joint torque sensors, detailed terrain geometry) that are not available on standard commercial platforms, hindering practical field deployment.

The core problem is the lack of a simple, reproducible, and sensor-independent energy model that accounts for the interaction between terrain slope and robot heading to enable energy-aware path planning.

2. Methodology

The authors propose a framework to identify a compact energy model using only standard onboard sensors (battery management system, IMU, and legged odometry).

A. Theoretical Framework

Coordinate System: The model defines the local slope angle ( $\alpha$ $α$ ) and the robot's heading relative to the slope direction ( $\gamma$ $γ$ ).
- $\gamma = 0^\circ$ : Facing directly uphill.
- $\gamma = 180^\circ$ : Facing directly downhill.
- $\gamma = 90^\circ$ : Moving perpendicular to the slope (lateral).
Energy Formulation: The total energy $E$ for a path is modeled as the time integral of the inner product between the generalized force (wrench) $f$ and the generalized velocity $\varpi$ :
$E = \int_{0}^{t_{goal}} f(\alpha, \gamma)^T \varpi(t) \, dt$
Key Assumption: The wrench $f$ is assumed to be a function solely of the slope angle $\alpha$ and heading $\gamma$ . This allows the problem to be reduced to identifying a low-dimensional function rather than a complex dynamic simulation.
Additivity: The model assumes energy is additive across trajectory segments, allowing path-level costs to be estimated by summing costs of local primitives.

B. Experimental Setup

Platform: Unitree B1 quadruped robot.
Sensors: Standard battery monitor (voltage/current), IMU (orientation), and legged odometry (pose/velocity). No external torque sensors or motion capture systems were used.
Environment: Outdoor tests on a controlled ramp and natural grassy terrain with slopes ranging from $5^\circ$ to $20^\circ$ .
Protocol: The robot executed straight segments at a constant speed ($0.3$ m/s) with various headings (uphill, downhill, cross-slope, and intermediate angles). Data was preprocessed to filter noise and ensure consistency between electrical power and mechanical velocity.

C. Calibration

The authors used a grid of tests to estimate the "force-equivalent cost" (the wrench components) for different combinations of $\alpha$ and $\gamma$ . By correlating battery power consumption with body-frame velocities, they derived empirical coefficients for the energy model.

3. Key Contributions

Sensor-Light Energy Model: A generic method to predict heading-dependent energy maps for quadrupeds using only ubiquitous onboard sensors (battery, IMU, odometry), avoiding the need for specialized torque instrumentation.
Heading-Dependent Analysis: A formal demonstration that energy cost is a function of both slope magnitude and heading direction, challenging the assumption that slope alone dictates energy expenditure.
Field-Validated Additivity: Experimental proof that energy consumption is additive across trajectory segments, validating the use of this model for global path planning (e.g., A* or Dijkstra algorithms with energy costs).
Practical Calibration Procedure: A reproducible protocol to identify the energy model parameters directly from field data.

4. Key Results

Force-Equivalent Cost vs. Slope:
- Uphill ( $\gamma=0^\circ$ ): Energy cost increases linearly with slope angle, as expected due to gravitational work.
- Downhill ( $\gamma=180^\circ$ ): Energy cost decreases with slope for forward motion, but not to zero. A significant portion of energy is still dissipated due to non-conservative effects (impacts, structural deformation, heat).
- Lateral Motion ( $\gamma=90^\circ$ ): Lateral movement remains energetically expensive regardless of slope direction and tends to worsen as the slope increases. This is attributed to the high effort required to maintain stability during cross-slope traversal.
Heading Asymmetry: The energy cost of moving downhill is not simply the inverse of uphill; the "gap" between forward and lateral costs highlights that heading is as critical as slope magnitude.
Rotational Cost: Preliminary results indicate that the torque required for pure rotation (yaw) increases with slope angle, suggesting that reorienting on an incline requires additional stabilization energy.
Superposition Validation: Experiments comparing composite paths (e.g., two segments) against direct paths showed a close match (within ~4-13% difference), confirming the model's assumption of linear energy accumulation.

5. Significance and Future Work

Significance:
This work bridges the gap between theoretical energy modeling and practical deployment for legged robots. By proving that a simple, sensor-based model can capture the complex interplay of slope and heading, the authors enable energy-aware motion planning. Planners can now select paths that are not just geometrically shortest but energetically optimal (e.g., choosing a longer, gentler slope over a shorter, steeper one, or avoiding lateral traverses on steep hills). This is crucial for extending mission duration in agriculture and disaster response.

Future Work:
The authors plan to:

Couple online slope estimation with real-time path-cost evaluation.
Broaden the model identification to include varying velocities, different gaits, diverse terrains, and different robot platforms to improve generalizability.
Refine the model to capture non-linearities and parameters currently unaccounted for in the simplified linear fit.