We propose a method for view-dependent simplification of articulated-body dynamics, which enables an automatic trade-off between visual precision and computational efficiency. We begin by discussing the problem of simplifying the simulation based on visual criteria, and show that it raises a number of challenging questions. We then focus on articulated-body dynamics simulation, and propose a semi-predictive approach which relies on a combination of exact, a priori error metrics computations, and visibility estimations. We suggest several variants of semi-predictive metrics based on hierarchical data structures and the use of graphics hardware, and discuss their relative merits in terms of computational efficiency and precision. Finally, we present several benchmarks and demonstrate how our view-dependent articulated-body dynamics method allows an animator (or a physics engine) to finely tune the visual quality and obtain potentially significant speedups during interactive or off-line simulations.