Agentic humanoids doing patrols

Humanoid patrol robots are moving from lab demos toward site-security tests, with Strike Robot showing a system that can spot people, track them, and keep moving around obstacles in real time. (arxiv.org, sotwe.com) A humanoid patrol system is a walking camera-and-sensor platform with software that decides what to inspect next, instead of waiting for a human operator to steer it. Strike Robot’s March 26, 2026 paper calls that stack “SafeGuard ASF” and says it runs on Unitree’s G1 humanoid robot. (arxiv.org, unitree.com) The paper says the robot combines RGB-D vision, which adds depth to ordinary video, with a ReAct-style reasoning system and learned walking controls. It targets three patrol jobs: fire or smoke detection, abnormal pipeline temperature checks, and intruder detection in restricted zones. (arxiv.org) Strike Robot’s recent demo posts show the same idea in a simpler scene: the robot identifies a designated person, follows that person, avoids obstacles, and replans its route as the environment changes. The company said the system uses face detection, distance estimation from camera data, A* path planning, and continuous replanning. (sotwe.com) In the paper, the team reported 94.2% mean average precision for fire and smoke detection with 127-millisecond latency. The authors also said they validated patrol, human detection, and obstacle avoidance in both simulation and real-world tests. (arxiv.org) The hardware matters because patrol robots need to fit through spaces built for people. Unitree markets the G1 as a compact humanoid with 23 to 43 joint motors and a starting price from $13,500, which is far below the price range long associated with custom humanoid research machines. (unitree.com) Factories and utilities have chased this kind of system because fixed cameras leave blind spots and separate sensors often cannot act on what they detect. Strike Robot’s paper argues a humanoid can carry its own sensors, move through stairs and narrow passages, and manipulate equipment if a patrol turns into an intervention. (arxiv.org) Other groups are pitching similar patrol use cases. Reports this month said Fangchenggang, on China’s border with Vietnam, is building a humanoid robot testing center with UBTECH Robotics for patrol and inspection work, showing that security and inspection are becoming an early commercial target for humanoids. (msn.com, earth.com) The limits are still plain in the source material. Strike Robot’s results are in a newly posted arXiv paper, not a peer-reviewed journal article, and the public demos are short clips rather than long-duration field deployments with failure rates, guard-handoff rules, or customer contracts. (arxiv.org, sotwe.com) What the videos and paper show, for now, is a narrower claim: a humanoid can patrol, notice a target, and make small navigation decisions on the move. The next test is whether that behavior holds up over full shifts in real industrial sites, where dust, glare, crowds, and false alarms decide what gets deployed. (arxiv.org, sotwe.com)