Multi-Task Learning of Active Fault-Tolerant Controller for Leg Failures in Quadruped robots

TL;DR

This work tackles the problem of maintaining quadruped locomotion under leg faults, including electrical power loss and joint locking. It introduces a hierarchical multi-task reinforcement learning controller with a shared front policy and task-specific low-level policies for health, power-loss, and locking states, aided by a fault discriminator and a reflection initialization technique. Training combines two stages with parallel environments, domain randomization, and a tailored reward curriculum to produce diverse, transferable gaits, achieving zero-shot Sim2Real transfer on the SOLO8 platform. Results show enhanced fault survivability, preserved planar mobility and velocity tracking, and successful real-world deployment, advancing proactive fault-tolerant control for outdoor quadrupeds.

Abstract

Electric quadruped robots used in outdoor exploration are susceptible to leg-related electrical or mechanical failures. Unexpected joint power loss and joint locking can immediately pose a falling threat. Typically, controllers lack the capability to actively sense the condition of their own joints and take proactive actions. Maintaining the original motion patterns could lead to disastrous consequences, as the controller may produce irrational output within a short period of time, further creating the risk of serious physical injuries. This paper presents a hierarchical fault-tolerant control scheme employing a multi-task training architecture capable of actively perceiving and overcoming two types of leg joint faults. The architecture simultaneously trains three joint task policies for health, power loss, and locking scenarios in parallel, introducing a symmetric reflection initialization technique to ensure rapid and stable gait skill transformations. Experiments demonstrate that the control scheme is robust in unexpected scenarios where a single leg experiences concurrent joint faults in two joints. Furthermore, the policy retains the robot's planar mobility, enabling rough velocity tracking. Finally, zero-shot Sim2Real transfer is achieved on the real-world SOLO8 robot, countering both electrical and mechanical failures.

Figures (9)

• Figure 1: Sim2Real tasks were conducted on SOLO8 robots under different leg statuses: (Top) Health, (Mid) Limit, and (Low) Weak. 'Limit' and 'Weak' denote scenarios where locking(we simulate with larger PD) and electrical faults occur in two joints within left-front leg, respectively. Our control scheme demonstrates diverse gait skills, adapting to leg joint faults while accurately tracking human commands.
• Figure 2: Scheme of the different controller components
• Figure 3: Architectural during the first training stage. ('Limit' denotes mechanical locking, 'Weak' denotes electrical power loss). Symmetric 'Health' states were reflected into 'Limit' and 'Weak' environments. This technique fosters the gaits transfer performance.
• Figure 4: Leg status randomly transforms during the second training stage. Three example episodes are given here.
• Figure 5: Joints data from 120-steps trajectories of 6 quadruped agents. (a) and (b) show the actions and states of Hip Joint and Knee Joint within fault-leg. (c) t-SNE visualization of the robot(all 8 joints) state space.