We studied path integration in dogs, Canis familiaris, required to return to the starting point of L-shaped outward paths in the absence of landmark information. The analysis of the distributions of errors in the return direction and distance showed that the dogs made both systematic and random errors. They tended to overestimate by about 6 degrees the amount of turning required at the end of the outward path to take the correct return direction. Random angular errors were characterized by a standard deviation of about 9 degrees. The dogs also tended to underestimate the correct return distance by about 6%, and random distance errors were characterized by a standard deviation of about 13% of the correct return distance. While random errors are assumed to stem simply from inaccuracies in kinaesthetic and vestibular systems, systematic errors are likely to be generated because path integration is implemented at the neural level through approximate rather than mathematically exact solutions. A path integration model in which the moving animal is assumed to overestimate slightly the inferred motion parallax of the starting point can roughly account for the dogs' directional bias, but not for systematic errors in distance estimation. Copyright 1998 The Association for the Study of Animal Behaviour. Copyright 1998 The Association for the Study of Animal Behaviour.