Scenario-based simulation testing has become a widely adopted approach for evaluating Autonomous Driving Systems (ADS) due to its scalability and flexibility. While existing testing methods have been effective in discovering safety-critical failures, they are largely driven by generic criticality metrics and provide limited support for controllable scenario generation toward specific testing objectives; in practice, test engineers often need to assess ADS behavior under targeted spatial–temporal interaction patterns, such as standardized regulatory scenarios. In this paper, we propose CONGRAD, a controllable scenario generation approach based on gradient-guided falsification. CONGRAD formulates target scenario generation as a falsification problem by transforming user-defined testing objectives into signal temporal logic (STL) specifications. To overcome the coarse-grained guidance of traditional robustness-guided falsification, CONGRAD exploits differentiable STL robustness semantics to compute robustness gradients with respect to scenario parameters, which are then used to guide scenario mutation in a fuzzing-based generation process. We evaluate CONGRAD on Baidu Apollo platform using representative testing specifications derived from the ISO~34502 standard. Experimental results demonstrate that CONGRAD can generate targeted scenarios more efficiently and with greater controllability than existing baselines, highlighting its effectiveness for controllable ADS testing.