Strike Robot launches SR Agentic

Strike Robot has launched SR Agentic, a software stack for humanoid robots that the company says is built for dangerous industrial inspections on Unitree’s G1 robot. (strikerobot.ai) Strike Robot’s public materials describe the company as a Singapore-based embodied artificial intelligence platform focused on “security and safety” work in nuclear plants, high-voltage facilities, and radiation zones. Its GitHub profile says its flagship system patrols, monitors, and intervenes in high-risk sites either autonomously or by teleoperation. (github.com) The hardware underneath is Unitree’s G1, a roughly 35-kilogram humanoid with 23 to 43 degrees of freedom, a depth camera, three-dimensional lidar, and about two hours of battery life. Unitree lists the base model at “from $13.5K” on its official product page. (unitree.com) Humanoid robotics is the effort to put software on a human-shaped machine so it can climb stairs, use tools, and reach equipment built for people. In factories and utility sites, that matters because valves, panels, ladders, and walkways were designed around human bodies, not wheeled robots. (unitree.com) The hard part is turning a language model into physical action without making the robot unstable. Strike Robot’s March 26 paper says its system combines camera-based perception, a ReAct-style reasoning loop, and learned walking policies so the robot can detect a problem, decide on a step, and move through the site. (arxiv.org) That paper, titled *SafeGuard ASF: SR Agentic Humanoid Robot System for Autonomous Industrial Safety*, describes three target jobs: fire and smoke detection, abnormal pipeline temperature checks, and intruder detection in restricted areas. The authors report 94.2 percent mean average precision for fire or smoke detection with 127-millisecond latency, plus real-world tests for patrol, human detection, and obstacle avoidance. (arxiv.org) Strike Robot is also building its own simulation layer around the G1. A repository published in early April describes a browser-based MuJoCo WebAssembly simulator with real-time joint control, keyframe recording, and support for loading pre-trained control policies before deployment on hardware. (github.com) That simulation-first approach matches a wider robotics pattern known as simulation-to-real transfer: train a policy in software, then move it onto the physical machine. Unitree’s own developer ecosystem also emphasizes sim-to-real workflows for the G1, which helps explain why newer humanoid startups are using the platform as a base instead of building a full robot from scratch. (github.com) (unitree.com) Strike Robot’s launch puts the company in the part of the humanoid market that is selling software and task-specific autonomy, not just a robot body. The test for SR Agentic now is whether those demos and early research results turn into repeatable deployments in the hazardous sites the company is targeting. (strikerobot.ai)