Auditable Feasibility-Gated Command Admission for Language-Conditioned Robotic Task Allocation

Abstract Language-conditioned robotic systems can map natural-language instructions to plausible skills, yet a semantically matched skill may still be non-executable on a particular robot. This paper presents QER–HFV, an auditable command-admission layer for robotic task allocation under semantic uncertainty. The layer expands the top-ranked skill hypotheses into explicit skill–robot candidates, checks capability, resource, communication, workspace, policy, and interface constraints, and records admission, rejection, or escalation outcomes with blocking reasons before any soft ranking is applied. In a reproducible decision-layer benchmark, 4-bit semantic memory preserves the FP32 allocation-success proxy to the displayed precision while reducing normalized semantic-memory cost to 12.5%. Hard feasibility verification lowers the single-step non-executable allocation rate from 0.580 for semantic-only retrieval to 0.108 for Semantic+HFV and 0.084 for the full gate. In a 1200-step online replay, the post-gate resource-aware ranker improves the mission-success proxy from 0.651 to 0.821, while windowed strong baselines clarify that the ranker is not a general batched scheduler. Execution-facing tests further show 100/108 executable successes and 72/72 correct infeasible-command rejections in a controlled tabletop Hiwonder xArm 1S study; overall correct handling is 172/180 (0.956; 95% Wilson CI [0.915, 0.977]). The simplified Robot Operating System (ROS)/Gazebo comparison records no collisions or infeasible false executions for QER–HFV. Targeted boundary experiments additionally show 0.9942 mean cosine preservation for 4-bit embeddings, extend candidate-fleet scaling to 80 robots per task, improve correct handling of 300 ambiguous/vague commands from 0.6267 to 0.8600 (paired difference +0.233; 95% bootstrap CI [0.167, 0.300]) with an escalation policy, and complete a 72-hour Raspberry Pi 5 replay with 259,200 decisions, two recoverable events, and no crashes or restarts. The results support QER–HFV as a reusable pre-command admissibility interface rather than a formal safety certificate or global scheduling optimizer.