Min Jung Kim

Minkyu Ham, Dongho Kim, Chan Lee et al.

Foundation models have given robots a sophisticated brain for complex decision-making, yet deploying that intelligence into a physical platform still demands tedious, expert-driven calibration. This deployment gap, the robot's spinal cord, remains a primary bottleneck to scalable Embodied AI. Hence, we propose SPINE (Scalable Physical Integration with ageNtic Expertise): an agentic framework for systematically debugging and deploying bimanual robots with minimal robotics expertise. SPINE's harness comprises two orchestrated multi-agent workflows: a profile builder that creates robot-specific context, and a debugger that cycles through diagnosis, repair, and validation until teleoperation works. Across seven DOBOT X-Trainer debugging scenarios, a robotics novice using SPINE outperformed human operators using Claude Code with the same reference materials, but without SPINE's structured workflow, improving operationalization success from 75% to 100% and reducing mean time-to-teleoperation from 16 min 45 s to 13 min 47 s. On AgileX PiPER, a distinct ROS/CAN bimanual arm, SPINE resolved all 10 implanted bugs, versus 9 out of 10 for the expert baseline, in nearly the same amount of time. Together, these results show that SPINE can transfer across bimanual platforms, reduce dependence on expert calibration, and move embodied AI closer to scalable real-world deployment.

Hyunkyung Han, Min Jung Kim

Background and Objective: Deep video models estimate left-ventricular ejection fraction (EF) from echocardiography with near-expert accuracy, and post-hoc attribution (Chefer relevance for transformers, Grad-CAM for CNNs) is increasingly used to certify that models "look at the right place." Yet whether these explanations are faithful both spatially and temporally is unaudited. Because EF is defined by the end-systolic (ES) and end-diastolic (ED) frames, a faithful explanation must localize the left ventricle (space) and the decisive frames (time). Methods: We fine-tune two distinct EF regressors on EchoNet-Dynamic -- a self-supervised VideoMAE transformer and a Kinetics-pretrained R(2+1)D CNN -- and audit each with architecture-matched attribution along three axes: intersection-over-relevance (IoR) against LV masks, deletion AUC, and a temporal localization index on ES/ED frames, each relative to chance with per-case 95% CIs over 50 studies. A tubelet-occlusion probe separates attribution failure from model behavior. Results: Both models are anatomically faithful -- IoR 2.91x (VideoMAE) and 1.98x (R(2+1)D) above chance -- yet temporally blind: temporal localization is indistinguishable from chance (0.97--1.00) and no better than random attribution. Occlusion shows the models do not preferentially rely on ES/ED (0.90x chance), so temporal blindness reflects model behavior, not an attribution artifact. Conclusions: Spatial faithfulness does not imply temporal faithfulness. Attribution can certify anatomical grounding while masking that a model ignores the clinically decisive frames -- a caution for XAI-based validation of video diagnostic models and a call for temporally-aware training and evaluation.