Design and Analysis of a Risk-Aware Reinforcement Learning Framework for Dynamic Contact Control in Fragile-Object Grasping

Abstract

Balancing active slip prevention and crush avoidance is challenging for fragile-object grasping. Conventional end-to-end systems often confound Out-of-Distribution (OOD) perceptual errors with underlying control deficiencies, obscuring true boundary robustness. This study proposes a risk-aware Reinforcement Learning (RL) framework and a decoupled evaluation architecture. A convolutional neural network-long short-term memory network validates temporal tactile slip representations offline, whereas the control loop is isolated from perceptual errors. A fragility sweep tightens the crush limit from 25.0 to 18.0 N using analytically computed proxy risk states and composite rewards to evaluate simulated policy boundaries. Policy training was assessed using three independent seeds, and each boundary condition was evaluated over 200 episodes. At 25.0 N, Sparse RL achieved a success rate of 83.5%. However, when the limit decreased to 19.0 N or lower, the same subset of trajectories collapsed into crush failures, yielding a 0% success rate and 83.5% crush-failure rate. Full RL retained a success rate of 62.0% at 18.0 N. Heuristic reflexes were conservative, while closed-loop tests with synthetically rendered tactile observations revealed severe performance degradation under OOD conditions. The decoupled architecture isolates control mechanics, demonstrating that unconstrained policies rely on transient high forces. In the simulated boundary tests, continuous risk optimization produces more stable contact than rigid heuristic strategies. Explicit risk constraints are critical for boundary control. The proposed framework provides a simulation-based paradigm for identifying control-degradation mechanisms, although real- robot validation remains a future work direction.