Abstract: In smart agriculture, the trend of autonomous navigation is increasingly prominent. However, the existing unstructured orchard ground is uneven, resulting in low path tracking accuracy for crawler-based picking robots. This study proposes an improved pure tracking navigation method based on fuzzy-immune PID control. Firstly, the Beidou satellite dual-antenna RTK device is used to real-time calculate the heading angle and the current position of the chassis; then, an improved pure tracking algorithm is applied to calculate the current heading deviation of the chassis, which is used as the control quantity to achieve path tracking; subsequently, since the classical PID control has problems such as difficult elimination of steady-state error and slow response, in this paper, based on the classical PID controller, combined with the immune feedback mechanism and fuzzy control for adaptive adjustment of the proportional parameters of PID, and further, using fuzzy control to correct the integral and differential parameters of PID, the stability and response speed of the navigation system are improved. Linear simulation shows that under the conditions of an initial angle of 30°, an initial lateral deviation of 0m, and a forward-looking distance of 0.5m, the online distance and online time of the algorithm proposed in this paper are 70cm and 2.4s, respectively, with an overshoot of 9.8cm; for comparison algorithm 1, the online distance and online time are 303cm and 10.3s, respectively, with an overshoot of 11.48cm; for comparison algorithm 2, the online distance and online time are 252cm and 8.5s, respectively, with an overshoot of 11.21cm; comparison algorithm 3 has not been online after traveling 10m. Under the conditions of an initial angle of 0°, an initial lateral deviation of 0.5m, and a forward-looking distance of 0.5m, the online distance and online time of the algorithm proposed in this paper are 81cm and 3.2s, respectively; for comparison algorithm 1, the online distance and online time are 488cm and 17.4s, respectively; for comparison algorithm 2, the online distance and online time are 235cm and 8s, respectively; comparison algorithm 3, the on-line time and on-line distance are 119cm and 4.3s. Under the conditions of an initial angle of 0°, an initial lateral deviation of 1m, and a forward-looking distance of 0.8m, the online distance and online time of comparison algorithm 1 are 563cm and 22.2s, respectively; comparison algorithm 2 has not been online after traveling 10m; comparison algorithm 3, the on-line time and on-line distance are 159cm and 6.4 s. Curve simulation shows that the mean lateral deviation of the improved algorithm experiments at three speeds is 2.12cm, 3.28cm, and 4.45cm, which is 50.2%, 41.8%, and 38.5% lower than that of the classic pure pursuit algorithm; the standard deviation is 0.02m, 0.0288m, and 0.0371m, which is 56.3%, 51.4%, and 49.7% lower than that of the classic pure pursuit algorithm; the root mean square error is 2.91cm, 4.37cm, and 5.79cm, which is 31.7%, 46.6%, and 43.9% lower than that of the classic pure pursuit algorithm. Field experiments show that when the speed is 0.3m·s⁻¹, the mean lateral deviation of the three experiments is 2.95cm, 3.04cm, and 3.44cm, and the root mean square error is 3.59cm, 3.71cm, and 4.34cm, and the standard deviation is 2.04cm, 2.13cm, and 2.64cm; when the speed is 0.4m·s⁻¹, the mean lateral deviation of the three experiments is 3.80cm, 4.75cm, and 3.53cm, and the root mean square error is 4.56cm, 5.80cm, and 6.35cm, and the standard deviation is 2.52cm, 3.33cm, and 4.15cm; when the speed is 0.5m·s⁻¹, the mean lateral deviation of the three experiments is 4.68cm, 4.97cm, and 5.67cm, and the root mean square error is 5.98cm, 6.16cm, and 6.98cm, and the standard deviation is 3.73cm, 3.65cm, and 4.07cm.