AI for Robotics Group

Overview

Embodied intelligence and reliable multi-robot execution

The AI for Robotics Group studies how heterogeneous robots can perceive their surroundings, reason about tasks and coordinate physical actions. Its research connects multimodal data, high-level planning, robot control and safety, with applications in last-100-metre delivery, campus and building inspection, and service assistance.

Task success is evaluated at execution: routes must satisfy graph constraints; actions must match each robot’s physical form and capabilities; handovers must maintain consistent timing and state; and the final state must meet the goal, such as delivering an item to its specified destination.

Research areas

• Multimodal perception and robotic data. Understanding objects and surrounding scenes through entity segmentation, 360-degree video segmentation and tracking, and the integration of language, vision, point clouds and robot-state data.
• Task reasoning and multi-robot coordination. Using language models and world models to decompose tasks, plan and optimize workflows, and coordinate mobile robots, manipulators and robot dogs. A coffee-delivery workflow links navigation, elevator use, grasping, handover and delivery.
• Embodied action and control. Studying vision-language-action models, reinforcement learning and manipulator control to translate plans into feasible motion and interaction, respecting the action limits of each robot platform.
• Safety and execution verification. Studying world-model safety and safe control, and checking route feasibility, robot capability constraints, handover consistency and final task completion throughout collaborative execution.
• Data-efficient skill learning and evaluation. SkillCost-Bench uses an initial set of 10 demonstrations per task and task features to predict model-specific learning curves, then allocates further data by expected marginal gain under a fixed budget. Its dual-arm Piper protocol includes 12 calibration tasks arranged as six single- and dual-arm pairs, plus four unseen tasks, evaluated through task success and process scores.