Presentation
Leveraging Transfer Learning to Enhance Generative Postural Variability for Ergonomics
SessionPoster Session 1
DescriptionGenerative AI offers promising advances for human factors and ergonomics, particularly in proactive risk identification and ergonomic intervention. However, limited variability in generative models remains a major challenge, especially when training data are scarce. This study investigates the use of transfer learning to enhance postural variability in generative lifting posture models. A conditional diffusion model based on a U-Net architecture was trained to generate lifting postures conditioned on hand position and body height. Transfer learning was applied by pre-training the model on common mid-range hand positions and fine-tuning it on rare extended-reach postures. Only the bottleneck and skip connections were fine-tuned to retain general feature extraction while adapting to new postural variations. Evaluation based on bootstrap sampling showed that transfer learning significantly improved postural variability without compromising similarity or validity when compared to models trained from scratch. Additionally, transfer learning greatly reduced training time, demonstrating improved computational efficiency. These results suggest that transfer learning can help generative models better capture natural movement variability, offering a practical solution to address data scarcity in ergonomic applications. This approach has the potential to support more personalized, diverse, and scalable posture prediction models for improving WMSD risk assessment and workplace ergonomics.
Event Type
Poster
TimeTuesday, October 14th5:30pm - 6:30pm CDT
LocationRiverside East
Similar Presentations