Aims. We describe the astrometric reduction of images obtained with the FORS2/VLT camera in the framework of an astrometric planet search around 20 M/L-transition dwarfs. We present the correction of systematic errors, the achieved astrometric performance, and a new astrometric catalogue containing the faint reference stars in 20 fields located close to the Galactic plane. Methods. Remote reference stars were used both to determine the astrometric trajectories of the nearby planet search targets and to identify and correct systematic errors. Results. We detected three types of systematic errors in the FORS2 astrometry: the relative motion of the camera's two CCD chips, errors that are correlated in space, and an error contribution of as yet unexplained origin. The relative CCD motion probably has a thermal origin and typically is 0.001-0.010 px (∼0.1-1 mas), but sometimes amounts to 0.02-0.05 px (3-6 mas). This instability and space-correlated errors are detected and mitigated using reference stars. The third component of unknown origin has an amplitude of 0.03-0.14 mas and is independent of the observing conditions. We find that a consecutive sequence of 32 images of a well-exposed star over 40 min at 0.6 seeing results in a median rms of the epoch residuals of 0.126 mas. Overall, the epoch residuals are distributed according to a normal law with a χ 2 value near unity. We compiled a catalogue of 12 000 stars with I-band magnitudes of 16-22 located in 20 fields, each covering ∼2 × 2 . It contains I-band magnitudes, ICRF positions with 40-70 mas precision, and relative proper motions and absolute trigonometric parallaxes with a precision of 0.1 mas/yr and 0.1 mas at the bright end, respectively. Conclusions. This work shows that an astrometric accuracy of ∼100 micro-arcseconds over two years can be achieved with a large optical telescope in a survey covering several targets and varying observing conditions.