DreamFLEX: Learning Fault-Aware Quadrupedal Locomotion Controller for Anomaly Situation in Rough Terrains

TL;DR

DreamFLEX targets fault tolerant quadrupedal locomotion under joint faults on rough terrains. It introduces DreamFLEX, a framework that combines an asymmetric actor-critic policy trained with proprioception and a Failure Estimation and Modulation Network (FEMNet) that explicitly estimates the joint fault vector $f_t$ and modulates the latent embedding $z_t$ to adapt gait in real time. The method uses a terrain curriculum and a joint fault curriculum during training and validates performance in both simulation (Isaac Gym) and real-world experiments on a Unitree Go1, outperforming prior fault tolerant methods and ablations. The results show robust locomotion despite faults, enabling traversal of rough terrains with reduced hardware damage and less need for human intervention, highlighting strong sim to real transfer.

Abstract

Recent advances in quadrupedal robots have demonstrated impressive agility and the ability to traverse diverse terrains. However, hardware issues, such as motor overheating or joint locking, may occur during long-distance walking or traversing through rough terrains leading to locomotion failures. Although several studies have proposed fault-tolerant control methods for quadrupedal robots, there are still challenges in traversing unstructured terrains. In this paper, we propose DreamFLEX, a robust fault-tolerant locomotion controller that enables a quadrupedal robot to traverse complex environments even under joint failure conditions. DreamFLEX integrates an explicit failure estimation and modulation network that jointly estimates the robot's joint fault vector and utilizes this information to adapt the locomotion pattern to faulty conditions in real-time, enabling quadrupedal robots to maintain stability and performance in rough terrains. Experimental results demonstrate that DreamFLEX outperforms existing methods in both simulation and real-world scenarios, effectively managing hardware failures while maintaining robust locomotion performance.

Figures (8)

• Figure 1: DreamFLEX enables the robot to autonomously detect leg joint faults during locomotion (with the rear right calf joint highlighted by a red circle). Subsequently, the policy is robustly adapted to overcome rough terrains, even under faulty conditions.
• Figure 2: Overview of our proposed training framework, DreamFLEX. The framework consists of three main components: (a) learning a quadrupedal locomotion controller based on asymmetric actor-critic architecture, (b) randomly assigned joint failure scenarios, and (c) a failure estimation and modulation network (FEMNet) for estimating the failure vector, $\boldsymbol{\mathrm{f}}_t$, and modulating the latent embedding, $\tilde{\boldsymbol{\mathrm{z}}}_t$, for the policy network.
• Figure 3: Examples of the fault scenarios in the front left calf joint of the quadrupedal robot: (a) locked joint and (b) weakened motor.
• Figure 4: The architecture of FEMNet. (a) FEMNet jointly estimates the body linear velocity $\boldsymbol{\mathrm{v}}_t$, joint fault vector $\boldsymbol{\mathrm{f}}_t$, and latent vector $\boldsymbol{\mathrm{z}}_t$, and (b) modulates the latent vector based on the joint fault vector.
• Figure 5: Learning curves of the (a) linear velocity tracking, (b) terrain level, and (c) total rewards. The higher the value, the better the performance. As an upper bound performance, the oracle policy was trained by directly accessing the privileged information. Except for the oracle policy, DreamFLEX the blue solid line, outperforms all the compared algorithms.