Risk-Aware Benchmarking and Adaptation for Gentle Retrieval Through Yielding Occlusions

Robots retrieving objects from cluttered environments often must interact with compliant structures that occlude a target while avoiding damage to both the surroundings and the robot. Existing manipulation benchmarks and policies typically emphasize task completion or aggregate interaction costs, making it difficult to isolate how occluder mechanics, retrieval stage, and rare contact events govern failure under yielding occlusions. We introduce GRAM, a benchmark for gentle retrieval through yielding occlusions that factorizes retrieval into Expose, Attach, and Transport stages, procedurally generates thin-shell and branching occluders with controlled morphology shifts, and defines success under explicit environment-disturbance and robot self-preservation budgets. Across modern visuomotor and vision-language-action policies, GRAM reveals three recurring failure modes: mechanics-driven distribution shifts alter effective interaction strategies, bottlenecks move across retrieval stages, and budget violations are often caused by short peak events rather than aggregate interaction cost. Motivated by these findings, we propose SRAH, a lightweight residual adapter that conditions a frozen policy on predicted retrieval stage, recent mechanics response, remaining budget headroom, and short-horizon hazard, while using unsafe and failed rollouts as counterexample supervision. Across ACT-style and OpenVLA backbones and both occluder families, SRAH improves budgeted gentle success by 4.1–5.7 percentage points under cross-family and tightened-budget evaluation while reducing environment and robot-side violations. Targeted real-robot analogues exhibit the same qualitative pattern, improving acceptable retrieval more than raw completion and suggesting that benchmark-derived structure can guide practical adaptation for damage-sensitive, contact-rich retrieval.