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

2026-06-14 · Source: Machine Learning · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

LoComposition introduces a novel approach to learning-based quadrupedal locomotion, decoupling control functions typically entangled in complex reward formulations. Instead, it employs distinct mechanisms: rewards for task specification, constraints for operational limits, energy minimization for gait preference, and exteroceptive perception for adapting energy use to terrain difficulty. This formulation eliminates explicit gait priors, such as air-time or foot-clearance targets, allowing gait behavior to emerge. Compared to conventional complex-reward baselines, LoComposition achieves comparable terrain traversal while significantly reducing the cost of transport by 56% and operational-limit violations by 96%. The resulting policies demonstrate zero-shot transfer to a physical Unitree Go2 robot using LiDAR-based elevation mapping.

Key takeaway

For Robotics Engineers developing autonomous quadrupedal systems, LoComposition offers a compelling alternative to traditional complex reward designs. By decoupling task, operational limits, energy minimization, and terrain adaptation, you can achieve significantly more efficient and robust locomotion. Consider adopting this modular control paradigm to reduce your robot's energy consumption by 56% and minimize operational violations by 96%, facilitating more reliable zero-shot deployment on diverse physical terrains.

Key insights

Decoupling quadruped locomotion control into distinct mechanisms enables efficient, terrain-adaptive behavior without explicit gait priors.

Principles

• Decouple task, limits, gait, and terrain.
• Energy minimization drives emergent gait.
• Exteroception adapts energy use to terrain.

Method

LoComposition uses distinct mechanisms: rewards for task, constraints for operational limits, energy minimization for gait preference, and exteroceptive perception for terrain adaptation, removing explicit gait priors.

In practice

• Reduce robot cost of transport by 56%.
• Minimize operational limit violations by 96%.
• Enable zero-shot transfer to physical robots.

Topics

• Quadrupedal Locomotion
• Reinforcement Learning
• Energy Efficiency
• Terrain Adaptation
• Robotics Control
• Unitree Go2

Best for: Research Scientist, Robotics Engineer, AI Scientist, Machine Learning Engineer

Related on AIssential

• Mind Your Steps: A General Learning Framework for Accurate Humanoid Foothold Tracking — The framework enables terrain-agnostic 3D humanoid foothold tracking by dynamically sampling goals and using a robust target representation for real-world transfer.
• Robots that refuse to fail: AI evolves 'legged metamachines' that reassemble and withstand injury — Modular robots with athletic intelligence can reconfigure, recover from injury, and maintain movement in dynamic environments.
• Metis: A Generalizable and Efficient World-Action Model for Autonomous Driving and Urban Navigation — Decoupling video generation from action prediction in WAMs, using a Mixture-of-Transformers and asymmetric attention, improves efficiency and generalization for autonomous navigation.
• Superhuman Safe and Agile Racing through Multi-Agent Reinforcement Learning — Multi-agent reinforcement learning with diverse self-play enables superhuman, safe, and generalizable high-speed robotic coordination in dynamic environments.
• TAM: Torque Adaptation Module for Robust Motion Transfer in Manipulation — TAM adapts torque commands at the robot interface, enabling robust motion transfer and decoupling policy training from robot dynamics.
• LLM-Foraging: Large Language Models for Decentralized Swarm Robot Foraging — Integrating LLMs at key decision points in swarm robotics enhances adaptability and performance without costly re-optimization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.