Robo-advisors provide investors with automated portfolio management services, and their growth is unprecedented. They performed less well during the COVID-19 pandemic than they had done during the non-crisis period. That less satisfactory performance was due to the fact that rare disasters are highly unlikely to occur and yet have a huge impact on financial markets. To improve the performance and learning efficiency of Robo-advising, this paper develops a novel learning framework to tackle rare disaster events. This framework integrates importance sampling with the reinforcement learning algorithm. Instead of sampling transition probability from a ground-truth probability distribution, we sample it from a proposal distribution, where the event of interest occurs more frequently. The proposed algorithm is validated by data covering both the 2008 financial crisis and the COVID-19 pandemic, showing superior performance over benchmarked methods. The estimated quarterly return of the robo-advisor portfolio using the proposed algorithm's optimal policy is 0.512%, significantly higher than both the benchmarked policy and the average quarterly return, which are -0.639% and -14.55% respectively. This is attributed to the intended learning about rare disasters so robo-advisors sharply lower the position on risky assets in advance. The proposed algorithm is model-free and reduces the variance of value estimates through importance sampling. In addition to methodological contributions, our study contributes to the growing literature on robo-advising by taking rare events into account. learn more on SSRN