Yes, padding will solve the issue. You have to choose a fixed size for time steps. Suppose if you choose fixed size 5, and some sample has just 3 values, then you have to pad it with some value (say 0 here), either before or after the data. If some sample has more than 5 values then you have to drop extra values from it, either from start or end. Remember not to lose much data.

This thing works, the network eventually learns to ignore the padding values.