The performance of vector-controlled induction motor drives without a speed sensor is generally poor at very low speed. The reasons are offset and drift components in the acquired feedback signals, voltage distortions caused by the nonlinear behavior of the switching converter, and the increased sensitivity against model parameter mismatch. New modeling and identification techniques are proposed to overcome these problems. A pure integrator is employed for stator flux estimation which permits high-estimation bandwidth. Compensation of the drift components is done by offset identification. The nonlinear voltage distortions are corrected by a self-adjusting inverter model. A further improvement is a novel method for online adaptation of the stator resistance. Experiments demonstrate smooth steady-state operation and high dynamic performance at extremely low speed.