Action recognition models often underperform on minority classes. Conventional augmentation, such as perturbing real data or generating photorealistic samples, enlarges dataset size but does not focus on mitigating class imbalance. To address class imbalance, we propose AutoPose, a pose-level augmentation framework that mixes real and synthetic poses (Real-Mix, Synthetic-Mix) and tunes interpolation weights via downstream validation to produce kinematically plausible, task-aligned samples without photorealistic rendering or extra real data. Across three datasets and two safety-critical classes (falling, cooking), AutoPose improves per-class Top-1 from baseline, e.g., +7.37% on Fall_Floor (HMDB51), +21.10% on Forward_Fall (GMDCSA), and +6.76% on Cook (Sims4Action). Rather than increasing dataset size indiscriminately, AutoPose maintains class-level semantic coherence by interpolating only between poses belonging to the same action category. The optimization over interpolation weights thus refines motion plausibility within this semantic boundary, ensuring generated samples enrich minority classes without distorting action meaning.