When a legged robot damages a leg, recovery algorithms typically rank candidate gaits by running short trials. But a gait that performs well briefly may not stay effective over longer stretches. The new work argues that this short-trial ranking is unreliable because it ignores how the robot's physical state evolves during continued execution.
The paper focuses on reset-free recovery, where the robot is not returned to a fixed starting pose between trials. In that setting, each trial changes the state that the next trial inherits, so the evaluation itself alters the conditions being evaluated. The authors propose a persistent profile-guided approach that tracks this state evolution and uses it to inform the final gait selection.
By incorporating the accumulated effects of prior trials, the method aims to choose a gait that remains viable over time, rather than one that merely excels in isolation. The work addresses a practical gap in damage-recovery algorithms for legged robots, where irreversible leg damage demands a durable solution, not just a promising short-term one.