Research Analyzer
← Back ICRA 2026

MULE � Multi-Terrain and Unknown Load Adaptation for Effective Quadrupedal Locomotion

Vamshi Kumar Kurva, Shishir Kolathaya

PDF

AI summary

Key figure (auto-extracted from paper)
An adaptive reinforcement learning framework enables quadrupedal robots to dynamically compensate for unknown payloads and traverse diverse terrains without predefined gaits or contact force sensors.
Reinforcement learning quadrupedal locomotion adaptive control payload adaptation legged robots sim-to-real transfer

Problem

Quadrupedal robots struggle to maintain stable locomotion under varying payloads and unstructured terrains, as traditional model-based controllers rely on predefined gaits, manual tuning, or offline parameter estimation that hinders real-time adaptability.

Approach

The method augments a nominal locomotion policy with a corrective adaptive policy that estimates foot forces and outputs real-time joint position adjustments to counteract payload shifts, trained in two phases via reinforcement learning.

Key results

  • Higher success rates and lower height-tracking errors across 2–10 kg payloads on flat ground, slopes, and stairs
  • Maintains energy efficiency comparable to no-load operation and top-performing baselines
  • Zero sim-to-real tuning required for successful real-world deployment on a Unitree Go1 with dynamic payloads
  • Adaptive policy dynamically modulates corrections to stabilize body height and velocity under shifting loads

Why it matters

Enables reliable, load-carrying legged robots for logistics and search-and-rescue applications without manual retuning or specialized hardware.

Abstract

Quadrupedal robots deployed for load-carrying ap- plications must maintain stable locomotion across diverse terrains and varying payloads. Traditional approaches like Model Predic- tive Control (MPC) can handle such variations but often rely on predefined gait schedules and manually tuned trajectory planners, limiting adaptability in unstructured environments. To address this, we propose an adaptive reinforcement learning (RL) frame- work that enables quadrupedal robots to respond dynamically to terrain and payload changes without relying on contact force measurements or gait designs. The controller consists of a nominal policy that learns general locomotion across terrains and an adap- tive policy that outputs corrective actions for handling dynamic variations due to payloads. We validate our approach through extensive simulations in Isaac Gym across payloads (2–10 kg) and terrains including flat ground, slopes, and stairs. Our method achieves higher success rates and lower height-tracking errors while maintaining the Cost of Transport (CoT) comparable to the best-performing baselines and to no-load (NL) operation. Real- world deployment on a Unitree Go1 confirms the approach’s effec- tiveness under both static and dynamic payload changes, including freely moving masses. The policy also performs well on outdoor terrains such as grass, soil, and staircases. The adaptive policy modulates corrections based on payload changes, improving body stability and tracking without post-deployment fine-tuning.

Index terms

Legged Robots Reinforcement Learning Robust/Adaptive Control

Related papers