Liquidity issues have been increasingly addressed recently, especially with regards to optimal execution of large orders. In practice, agents facing these issues are uncertain about the future dynamics of the assets, and face the risk of model misspecification, which could make the execution strategies non optimal. In this paper, we address the problem of uncertainty faced by an agent wishing to execute large orders on multiple assets. The agent only has knowledge about the distribution of the future drift of the assets composing her portfolio. We build on the work in [13] who proposed a model coupling Bayesian learning and dynamic programming techniques. More precisely, in this article, we provide a rigorous solution to the problem of portfolio optimal execution where prices have drifted Bachelier dynamics with an unknown drift. The agent uses Bayesian learning to update her estimation of the drift, while she maximizes the expected exponential utility of her final wealth. We consider the specific case where the prior is a non-degenerate multivariate Gaussian, and the costs are quadratic. We use stochastic optimal control tools to show how the problem of optimal execution simplifies into a system of ordinary differential equations (ODEs) which involves a matrix Riccati ODE with time-dependent coefficients for which classical existence theorems do not apply. However, using a method similar to the one in [10], we provide a rigorous solution to the problem by using a priori estimates obtained thanks to the original control problem.