This paper deals with deterministic and stochastic (risk sensitive) nonlinear filters. In the first part, the following deterministic filter model is considered: state equation, \(\dot x_t=f(x_t)+ \sigma(x_t) w_t\), and (accumulated) observation, \(\dot y_t=h(x_t)+v_t\), where \(w_t\) and \(v_t\) are deterministic disturbances. Applying the mean square disturbance error criterion, the authors formulate a robust filter with level \(\mu\) and study the minimax filter, which turns out to be robust for an appropriate range of \(\mu\). In the second part, the following stochastic model is treated: state equation, \(dX_t=f(X_t)dt+\sqrt \varepsilon \sigma (X_t)dB_t\), and observation, \(dY_t=h(X_t)dt +\sqrt\varepsilon d\overline B_t\), where \(B\) and \(\overline B\) are independent Brownian motions. The authors consider the risk sensitive filter for \(X_T\), which minimizes the exponential criterion, \(E\exp({\mu \over\varepsilon} |X_T-e_T|^2)\), by choosing suitable \(\{Y_s,s\leq T\}\) measurable \(e_T\), and show that a risk sensitive filter tends to the minimax filter of the first part as \(\varepsilon \to 0\).