PhysGuided-PPO: Physics-guided proximal policy optimization for energy-dissipative soft grasping

Dynamic soft grasping involves short-duration impacts, rapid energy transfer, and transient contact-pressure peaks that are difficult to regulate using simplified controllers or data-driven exploration alone. This paper proposes PhysGuided-PPO, a lightweight physics-guided extension of proximal policy optimization for energy-dissipative pneumatic soft grasping. An offline-trained one-step predictor estimates the immediate impact response of candidate valve commands and uses these predictions to bias the Bernoulli logit of the on-off valve and the Gaussian mean of the proportional valve during action sampling. Simulations over 0.18-0.50 m/s impacts show that PhysGuided-PPO increases energy dissipation from 91.2% and 93.4% to 96.8%, while reducing peak contact pressure from 2.46 and 2.32 N/cm² to 2.17 N/cm². These results indicate that short-horizon predictive guidance can improve impact-aware valve control while retaining the on-policy PPO training structure.