PAIRS Lab

Preprint 2026

SafeGuard ASF

SR Agentic humanoid robot system for autonomous industrial safety

Thanh Nguyen Canh, Thang Tran Viet, Thanh Tuan Tran, Ben Wei Lim

Strike Robotics · PAIRS Lab

arXiv Code Demos
A Unitree G1 recognizes a worker in protective gear and keeps tracking them across the factory floor.

01 · Abstract

A humanoid that patrols the dark factory

Unmanned "dark factories" run without people on site, so they need safety systems that can detect and respond to several kinds of hazard on their own. Fixed cameras and sensors leave blind spots and cannot go and check.

SafeGuard ASF (Agentic Security Fleet) deploys humanoid robots for autonomous hazard detection. It combines RGB-D perception, a ReAct-based agent that reasons and calls tools through the ToolOrchestra framework, and learned locomotion policies on the Unitree G1. It handles three hazards: fire and smoke, abnormal pipeline temperature, and intruders in restricted zones.

Locomotion policies, including motion tracking and velocity control, are trained with PPO in Unitree RL Lab and validated sim-to-sim before deployment. The system is tested in Isaac Sim and on a real G1: autonomous patrol, human detection and tracking, obstacle avoidance and approaching a person.

02 · Demos

SR Agentic on the Unitree G1

Real-robot and simulation runs of the SR Agentic stack, from the Strike Robot team. Each clip links to its original post.

Notice, reason, decide, actAn obstacle appears mid-mission; the robot notices it, reasons about it and reroutes, with its agent log on screen. May 2026 ↗
Tracking a workerThe robot recognizes a worker in protective gear and keeps tracking them across the floor. Apr 2026 ↗
Goal-driven object searchGiven a task such as "find the white bottle on the table", the robot searches and localizes the object in 3D. Apr 2026 ↗
What the robot seesThe onboard view: face detection of the target person, real-time obstacle detection and 3D positions of objects. Mar 2026 ↗
One line of text, in simulation"Pick and place the cylinder into the basket": a G1 in a factory simulation carries out a plain-language command. Mar 2026 ↗
Learned whole-body motionA fast martial-arts warm-up, showing the agility of the trained motion policies. Apr 2026 ↗
SR Agentic overviewThe plug-and-play agentic framework, starting with the Unitree G1, for autonomous anomaly detection. Apr 2026 ↗

03 · System

Six layers from sensors to actions

Six-layer architecture: RGB-D, thermal and IMU feed a perception layer (fire/smoke, thermal, depth, person), then understanding (severity, trend, fusion), memory (facility map, baselines, personnel), decision with ReAct reasoning and 23 tools, planning (D* Lite, MPPI) and action (locomotion, communication, actuation) on the Unitree G1.
Fig. 1: the six-layer architecture of SafeGuard ASF.
  1. L1

    Perception

    YOLOv8 fire and smoke detection, thermal monitoring, depth localization and person detection with Re-ID.

  2. L2

    Understanding

    Fuses detections to estimate hazard severity and predict how it evolves.

  3. L3

    Memory

    Facility map, thermal baselines per inspection point, and a database of authorized personnel.

  4. L4

    Decision: ToolOrchestra

    A ReAct agent loops thought, tool call and observation over perception, reasoning, knowledge and actuation tools.

  5. L5

    Planning

    D* Lite global waypoints and MPPI local trajectory optimization.

  6. L6

    Action

    Velocity-command locomotion policies, alerts to the control center and spoken warnings.

Locomotion training pipeline: Isaac Sim with 4096 environments and domain randomization, PPO with RSL-RL, policy validated in MuJoCo sim-to-sim and deployed sim-to-real on the Unitree G1.
Fig. 2: locomotion policies are trained in Isaac Sim with 4096 parallel environments and domain randomization, checked in MuJoCo, then deployed on the G1.

04 · Hazard scenarios

Fire, overheating pipes and intruders

1

Fire and smoke

Detect with a fine-tuned YOLOv8-m, estimate severity, localize in 3D, alert and trigger suppression, then monitor until it is out.

2

Abnormal pipeline temperature

Compare thermal readings with the stored baseline, follow the temperature trend, and trace the hot spot to its root cause, such as a stuck valve.

3

Intruder in a restricted zone

Detect people with YOLOv8-n and check them against authorized personnel by OSNet re-identification before raising an alert.

Fire response timeline: detection at T=0 (YOLOv8, fire 92%), localization at T+3 s (depth 6.2 m), alert at T+18 s with power off, monitoring until the fire reads 0% at T+48 s.
Fig. 3: a fire response from detection to confirmed extinguishment.

05 · Results

Detection, response time and compute

End-to-end response time by phaseseconds · Table III · hover a row for the breakdown

Latency per module on NVIDIA Jetson Orinms · lower is faster · Table VI · the ReAct reasoning runs on the CPU at about 0.85 s per decision, asynchronously

Limitations reported in the paper: few real-world scenarios so far, a single robot without fleet coordination, no physical manipulation (valves and suppression are actuated remotely), and edge-compute limits. In simulation, most failures came from missed detections in dispersed smoke.

06 · Simulation and real world

From Isaac Sim to a real G1

Isaac Sim factory: the humanoid on an automatic patrol route.
Fig. 7b: automatic patrol in an Isaac Sim factory.
Isaac Sim factory: the humanoid identifying a worker.
Fig. 7c: security and worker identification in simulation.
Real indoor environment, the robot's visual perception with bounding boxes and 3D positions, and a bird's-eye view map.
Fig. 9a–c: the real environment, visual perception with 3D localization, and the bird's-eye view.
The real G1 avoiding an obstacle, reaching a person and shaking hands.
Fig. 9d: the real G1 avoids an obstacle, reaches the person and shakes hands.

07 · Cite

Citation

@misc{canh2026safeguard,
  title         = {SafeGuard ASF: SR Agentic Humanoid Robot System for Autonomous Industrial Safety},
  author        = {Canh, Thanh Nguyen and Tran Viet, Thang and Tran, Thanh Tuan and Lim, Ben Wei},
  eprint        = {2603.25353},
  archivePrefix = {arXiv},
  year          = {2026}
}