Reliability-Aware Multimodal Attention and Adaptive Feedback in Low-Cost Social Robotics: A Five-Seed Perception-to-Action Evaluation

This preprint reports a negative result from a controlled simulation study of low-cost social robots that must act on visual, pose, state, language, and temporal observations that may be degraded or stale. It evaluates two mechanisms in a physically actuated PyBullet differential-drive environment: (1) MDTA, a reliability-biased joint multimodal attention block that adds a learned reliability log-bias to attention logits, and (2) RA-AFL, a reliability-aware adaptive feedback learning procedure. The study uses a 2 × 2 × 2 factorial ablation over log-polar visual tokenization, MDTA, and RA-AFL, plus a matched Cartesian-48 control. It spans five independent training seeds, 1,080 episodes (51,832 transitions), eight sensor-corruption conditions, and 90 paired out-of-distribution closed-loop episodes per policy. Key findings: Adding MDTA to log-polar tokenization lowered OOD intention accuracy from 94.72% to 86.46%. The full configuration reached 84.59% and raised expert-action MAE to 0.1112. In closed-loop evaluation, the full policy achieved 34.44% task success versus 91.11% for the privileged simulation expert, while travelling only 0.0396 m on average versus 0.9083 m. Its zero observed collision rate reflects a near-stationary policy, not safe navigation. Code inspection identifies a mechanism: RA-AFL trains a zero-action feedback target when minimum modality reliability falls below 0.38, and runtime gating forces zero action when trust falls below 0.28, which encourages stopping under uncertainty. Reliability calibration was weak, with some Spearman correlations near zero. The state target was constant in the severe-condition subset, so its correlation is not estimable. The paper does not claim performance superiority. It provides a transparent failure analysis and a redesign for the feedback objective: a task-conditioned safe-action set, a progress term, independently supervised corruption, abstention metrics, and a verified safety filter. All results are from simulation with scripted agents and a privileged simulator expert. No real-robot, human-subject, or sim-to-real claims are made.