Robots Are Leaving the Lab

How unstructured work is reshaping autonomy

Date: January 12, 2025
An Ursa Cortex Blog by Akash Iyer

Jobs like inspection, construction, and other unstructured tasks are pushing robotics to become far more robust. Unlike controlled factory floors, these environments are dynamic, messy, and unpredictable. This week’s post explores how robots are being developed to handle that complexity, from agentic AI systems used in industrial inspections to quadruped robots navigating active construction sites.

Spot and Agentic AI for End-to-End Industrial Inspections

Source: Institution of Mechanical Engineers

Boston Dynamics’ Spot is being paired with IFS industrial AI to create an end-to-end autonomous inspection loop. In this setup, Spot patrols industrial sites such as factories and power plants, gathers inspection data, and feeds that information into an AI system that analyzes findings and triggers follow-up actions.

The goal is to help asset-heavy industries manage labor and skills shortages, reduce human exposure to hazardous environments, and shift maintenance strategies from reactive to predictive.

What stands out is the emphasis on closing the loop beyond simply collecting images. The article highlights Spot’s ability to handle messy, real-world sensing tasks, including detecting thermal anomalies, identifying leaks, reading gauges, checking indicators, and spotting hazards. The agentic layer then turns these observations into prioritized actions, such as dispatching crews, scheduling maintenance, or flagging likely failures. In this framing, robots become part of an operational workflow rather than a standalone gadget.

Quadruped Robots Inspecting Construction Progress with BIM and Foundation Models

Source: Automation in Construction (Jan 2026)

This research tackles a classic challenge in construction robotics. Job sites change constantly, and digital Building Information Models rarely match reality perfectly. That mismatch often breaks autonomous systems.

The authors propose a framework where a quadruped robot uses BIM models to initialize planning but continuously verifies the environment using sensor data. The system maintains a map that distinguishes between fixed structures, missing elements, and dynamic obstacles, allowing navigation to remain robust even as the site evolves.

Inspection is made more general through adaptive next-best-view planning, where the robot decides what to inspect next, and the use of vision–language models for object inspection without task-specific training. The approach is validated in both simulation and real-world tests. The broader takeaway is that field robots are increasingly combining strong mobility, model-based priors like digital twins, and foundation-model perception to survive real, changing environments.

Fun Fact

Spot’s inspection toolbox can include thermal checks for overheating, listening for air or gas leaks, reading analog gauges, checking indicator lights, and spotting hazards like spills. These are the same routine sensing tasks humans perform during manual inspection rounds.

Published in Ursa Cortex: The Ursa Majors Group Blog