Reliable, Data-Efficient Autonomy: Methods and Deployment from Manufacturing Quality Monitoring to Robotic Manipulation

Across domains, autonomy spans decision support, shared autonomy, and full autonomy. Regardless of autonomy level, deployed systems must be data-efficient, reliable under shift, and compatible with human goals and constraints. This talk presents a principled, deployment-grounded framework for reliable, data-efficient autonomy across two equally central application areas in my group, manufacturing quality monitoring and robotic manipulation, organized around a three-thrust architecture. Each thrust includes a set of intervention points spanning feature/representation design, model/structure design, learning and adaptation strategies, objective/reward design, and runtime guardrails. Thrust 1: Transfer and adaptation with limited real data. Methods for reusing prior experience across machines, materials, tasks, and operating conditions via uncertainty-aware adaptation, targeted data acquisition, and cross-domain transfer. Thrust 2: Hard structure for reliability and safety. Physics- and constraint-aware approaches introduced through physics-informed representations, structured designs, physics-aware learning strategies and objectives, and safety monitoring mechanisms. Thrust 3: Soft structure for human value alignment. Mechanisms for incorporating intent, preferences, and acceptable tradeoffs through shared-autonomy interfaces, preference/correction signals, and objective/constraint design. The framework is illustrated with two case studies. In a manufacturing quality monitoring setting, physics-informed feature representations support reliability under shift with limited labels, active transfer learning enables across-machine knowledge transfer, and explainable quality assessment helps close the loop between model predictions and practitioner trust. In a robotic manipulation setting, stable in-hand object manipulation is achieved via physics-informed multi-agent reinforcement learning, paired with simulation-to-real transfer, and intent-driven reflective control that adapts behavior to user goals and context. Together, these case studies demonstrate a general blueprint for systems that are reliable, data-efficient, and human-compatible in deployment.

Dr. Xiaoli Zhang is an Associate Professor at the Colorado School of Mines, where she directs the Intelligent Robotics and Systems Laboratory and serves as Robotics Program Director. Her research develops AI-enabled modeling, control, and decision-making methods that are data-efficient, reliable under uncertainty and shift, and deployable in real systems, with applications in advanced manufacturing (metal additive manufacturing), energy systems, materials discovery, and robotics. Her work has been supported by NSF, ONR, USAF, DOE, DoD, and industry partners, and she received a 2017 NSF CAREER Award. She received her MS in Mechatronics and Automation Engineering from Xi’an Jiaotong University (2006) and her PhD from University of Nebraska Lincoln (2009). She serves as an AI/automation lead for the Alliance for the Development of Additive Processing Technologies (ADAPT) industrial-academia consortium and is an Associate Editor for IEEE Transactions on Industrial Informatics and IEEE Robotics and Automation Letters.