Soft fault compensation plays an important role in mobile robot locating, mapping, and navigating. It is difficult to achieve fast and accurate compensation for mobile robots because they are usually highly non-linear, non-Gaussian systems with lim-ited computation and memory resources. An adaptive particle filter is presented to compensate two kinds of soft faults for mobile robots, i.e., noise or factor faults of dead reckoning sensors and slippage of wheels. Firstly, the kinematics models and the fault models are discussed, and five kinds of residual features are extracted to detect soft faults. Secondly, an adaptive particle filter is designed for fault com-pensation, and two kinds of adaptive scheme are discussed: 1) the noise variances of linear speed and yaw rate are adjusted according to residual features; 2) the particle number is adapted according to Kullback-Leibler divergence (KLD) of two approximate distribution denoted with two particle sets with different particles, i.e., increasing particle number if the KLD is large and decreasing particle number if the KLD is small. The theoretic proof is given and experimental results show the effi-ciency and accuracy of the presented approach. Soft fault compensation plays an important role in mobile robot locating, mapping, and navigating. It is difficult to achieve fast and accurate compensation for mobile robots because they are usually highly non-linear, non-Gaussian systems with lim-ited computation and memory resources . First, the kinematics models and the fault models are discussed, and five kinds of residual features are extracted to detect soft faults. 2, an adaptive particle filter is designed for fault com-pensation, and two kinds of adaptive schemes are discussed: 1) the noise variances of linear speed and yaw rate are adjusted according to residual features; 2 ) the particle number is adapted according to Kullback-Leibler divergence (KLD) of two approximate distribution pronounced with two particles sets with different particles, ie increasing particle number if the KLD is large and decreasing particle number if the KLD is small. The theoretic proof is given and experimental results show the effi-ciency and accuracy of the presented approach.