🤖 machine learning

LoComposition: Terrain-Adaptive Energy-Efficient Quadruped Locomotion without Gait Priors

LoComposition introduces a learning-based framework for quadruped locomotion that decouples task specification, operational limits, gait preference, and terrain adaptation into distinct mechanisms, enabling zero-shot transfer to physical robots with significantly improved energy efficiency and reduced constraint violations compared to conventional complex-reward baselines.

Original authors: Loukas Kordos, Leonard T. Franz, Simon Rappenecker, Oliver Hausdoerfer, Angela P. Schoellig, Pavel Kolev, Georg Martius

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Loukas Kordos, Leonard T. Franz, Simon Rappenecker, Oliver Hausdoerfer, Angela P. Schoellig, Pavel Kolev, Georg Martius

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 teaching a four-legged robot dog how to run through a rocky, uneven forest. For a long time, engineers tried to teach this by giving the robot a massive, complicated list of rules: "Lift your left foot exactly 5 inches high," "Keep your legs in the air for 0.3 seconds," "Don't touch the ground too hard," and "Run like a horse." This is like trying to teach a human to dance by forcing them to count every single step and hold their arms in a specific shape. It works, but it's rigid, energy-draining, and if the floor changes, the robot gets confused.

The paper "LoComposition" proposes a smarter, more natural way to teach the robot. Instead of micromanaging how the robot moves, they teach it what to do and let the robot figure out the best way to do it.

Here is how they did it, broken down into simple concepts:

1. The Old Way: The "Strict Choreographer"

Previous methods used a single, complex "scorecard" (reward function) that tried to control everything at once. It told the robot exactly how to move its legs (gait priors), how to stay safe, and how to follow commands.

The Problem: It was like a strict choreographer who wouldn't let the dancer improvise. If the terrain changed, the robot kept trying to follow the rigid steps, wasting energy and sometimes breaking its own joints because it was trying too hard to fit a square peg in a round hole.

2. The New Way: The "Smart Coach" (LoComposition)

The authors split the teaching process into four distinct roles, like a sports team where everyone has a specific job:

The Goal Setter (Task Specification): This part simply says, "Go that way at that speed." It doesn't care how you get there, just that you get there.
The Safety Guard (Operational Limits): This acts like a bouncer at a club. It doesn't tell the robot how to dance; it just says, "If you twist your knee too hard or spin too fast, you're out." It sets hard boundaries to keep the robot from breaking itself.
The Budget Manager (Energy Minimization): This is the secret sauce. Instead of saying "Trot like a horse," the coach says, "Try to use as little battery as possible." The robot is smart; it quickly realizes that a specific type of run (a "trot") uses less energy than a bounding jump. So, it naturally chooses the efficient trot without being told to do so.
The Eyes (Perception): This is the robot's LiDAR (laser eyes) that scans the ground ahead. It tells the robot, "Hey, there's a rock coming up." This allows the robot to spend a little extra energy only when it needs to step over a rock, and save energy when the path is flat.

3. The Results: A Natural, Efficient Runner

When they tested this new method on a real robot (the Unitree Go2), the results were impressive:

No "Gymnastics" Required: They removed all the rules about how high to lift legs or how long to stay in the air. The robot figured out the best way to move on its own.
Massive Energy Savings: The robot used 56% less energy to move compared to the old, rule-heavy methods. It was like switching from a gas-guzzling truck to a hybrid car.
Fewer Breakages: The robot broke its "safety limits" (like twisting a joint too far) 96% less often. The "Safety Guard" kept it out of trouble.
Zero-Shot Transfer: This is the "magic trick." They trained the robot in a computer simulation, and when they put it on a real robot with real rocks, it worked immediately. They didn't have to retrain it or tweak the settings. It just worked.

The Big Picture Analogy

Think of the old method as teaching a child to walk by giving them a script: "Step left, lift foot 2 inches, wait 1 second, step right." If they trip on a pebble, they freeze because the script didn't say what to do.

The LoComposition method is like teaching a child to walk by saying: "Walk to the tree," "Don't hurt your knees," "Try not to get tired," and "Look where you are going." The child naturally figures out the most efficient way to walk, jumps over the pebble when they see it, and keeps their knees safe, all without a script.

In short: The paper proves that you don't need to hard-code specific walking styles (gaits) into a robot. If you give it a clear goal, safety limits, a reason to save energy, and eyes to see the ground, it will naturally learn to walk efficiently and safely on rough terrain.

Technical Summary: LoComposition

Problem Statement

Learning-based quadrupedal locomotion has achieved significant success in sim-to-real transfer and rough-terrain traversal. However, standard approaches typically rely on a single, complex reward function that entangles conceptually distinct objectives: task specification (velocity tracking), operational limits (actuator safety), gait preferences (e.g., specific contact timing or air-time), and terrain adaptation.

The authors argue that encoding these distinct functions into a single hand-weighted reward formulation leads to suboptimal learning outcomes. Specifically, explicit gait-style priors (such as desired air-time, contact counts, or foot clearance targets) act as engineering shortcuts that bias the learned motion toward specific patterns, potentially preventing the emergence of more efficient gaits and complicating the reward engineering process.

Methodology: LoComposition

The paper proposes LoComposition, a learning formulation that decomposes locomotion into four distinct mechanisms, each handled by a specific methodological component rather than a monolithic reward. The authors refer to their core components as LEP (Limits, Energy, Perception).

Task Specification (Velocity Tracking):
The task is defined strictly as tracking operator-specified planar and yaw velocities ( $v_{cmd}$ ). This is handled via a standard reward function ( $r_{track}$ ) that encourages alignment with velocity targets using Gaussian kernels.
Operational Limits (Constraints):
Instead of penalizing violations of hardware limits (torque, joint velocity, acceleration, etc.) within the reward, the authors use Constraints-as-Terminations (CaT).
- Violations of scalar limits ( $q_i \leq q_{max}$ ) are mapped to a stochastic termination probability ( $\delta_t$ ).
- If a limit is violated, the rollout is terminated with a probability proportional to the violation magnitude.
- This scales the per-step reward and reduces the continuation factor, effectively penalizing unsafe behavior without requiring dense reward shaping for every limit.
- Crucially, constraints are reserved only for physical/operational limits, not for gait style.
Gait Preference (Energy Minimization):
Rather than prescribing a specific gait (e.g., trotting or bounding) via reward terms, the formulation minimizes mechanical energy consumption.
- A penalty term ( $\ell_{Power}$ ) is added to the reward, proportional to the sum of absolute mechanical power ( $\sum |\tau_j \dot{q}_j|$ ) across all joints.
- This term acts as a preference signal among feasible motions, encouraging the emergence of low-cost-of-transport (COT) gaits without explicitly defining footfall patterns.
Terrain Adaptation (Exteroceptive Perception):
To enable predictive energy minimization on rough terrain, the policy receives exteroceptive observations.
- The input includes a body-centric $13 \times 11$ terrain elevation map (8 cm resolution) derived from LiDAR.
- This allows the policy to selectively spend energy (e.g., increasing foot clearance) only when the terrain requires it, resolving the tension between energy efficiency and obstacle traversal.

Training Setup: Policies are trained using Proximal Policy Optimization (PPO) in Isaac Lab on the Unitree Go2 robot. The training curriculum includes procedurally generated uneven terrain, stairs, and slopes.

Key Contributions

Decomposition of Locomotion: The paper presents a formulation that separates task tracking, operational limits, energy minimization, and perception into distinct mechanisms, removing the need for a single complex reward function.
Emergent Gait without Priors: The authors demonstrate that explicit gait priors (air-time, contact count, foot clearance) are unnecessary. Efficient, terrain-adaptive locomotion emerges naturally when the learning problem is decomposed as proposed.
Zero-Shot Sim-to-Real Transfer: The resulting policies transfer zero-shot to a physical Unitree Go2 using LiDAR-based elevation mapping, without retraining or privileged terrain information.

Experimental Results and Ablations

The authors validate their approach through ablation studies comparing their formulation (LEP) against a baseline using a complex reward function (RP) and various component removals.

Efficiency: Compared to the complex reward baseline (RP), LoComposition (LEP) reduces the Cost of Transport (COT) by 56% and operational-limit violations by 96%.
Ablation of Gait Priors (H1): Removing gait-prior constraints (comparing LEP vs. LCEP/LCP) shows that encoding priors as constraints harms terrain traversability and prevents the emergence of the most efficient gait (trotting).
Role of Energy Minimization (H2): Removing energy minimization (LP) results in a policy that traverses terrain but settles on an energetically expensive "bounding" gait. Adding energy minimization shifts the behavior to a lower-COT trotting pattern without explicitly prescribing it.
Role of Perception (H3): Removing perception (LE) leads to a policy that fails to adapt energy use to terrain difficulty, resulting in poor terrain traversal levels. Perception allows the robot to maintain low clearance on flat ground while selectively increasing it for obstacles.
Necessity of Constraints (H4): Removing operational-limit constraints (EP) allows the policy to exploit simulator inaccuracies, leading to severe violations of hardware limits (53.4% violation rate vs. 0.5% for LEP), rendering the policy undeployable.
Real-World Performance: In zero-shot hardware tests, LEP successfully traverses complex obstacle sequences (ramps, platforms) where the non-energy-minimized variant (LP) fails or violates limits. LEP achieves a 56% improvement in energy efficiency on hardware compared to the baseline.

Significance and Claims

The paper claims that explicit gait-style priors are unnecessary for rough-terrain quadrupedal locomotion. By separating operational limits (handled via constraints) from gait preferences (handled via energy minimization), the system allows efficient, natural gaits to emerge.

The authors position this as a shift in formulation principle: locomotion learning should be organized around the requirements and preferences the robot must reconcile (safety, task, efficiency, perception) rather than the specific gait pattern it should follow. This approach yields robust, energy-efficient, and deployable policies that outperform complex reward baselines in both simulation and real-world deployment.

Limitations: The authors note that their results are specific to the Unitree Go2 and the tested terrain distributions. They do not claim this is the only possible decomposition, nor do they address sample efficiency improvements from off-policy methods or formal runtime safety guarantees beyond the stochastic termination approach used.