💻 computer science

SafeGuard ASF: SR Agentic Humanoid Robot System for Autonomous Industrial Safety

This paper introduces SafeGuard ASF, a comprehensive framework that deploys Unitree G1 humanoid robots equipped with multi-modal perception, ReAct-based agentic reasoning, and PPO-trained locomotion policies to autonomously detect and respond to industrial hazards such as fire, temperature anomalies, and intruders in unmanned "dark factories."

Original authors: Thanh Nguyen Canh, Thang Tran Viet, Thanh Tuan Tran, Ben Wei Lim

Published 2026-03-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Thanh Nguyen Canh, Thang Tran Viet, Thanh Tuan Tran, Ben Wei Lim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a factory that never sleeps. It's a "dark factory," meaning it runs 24/7 with no humans inside, just machines humming along. The problem? If a pipe starts leaking hot steam, a machine catches fire, or a stranger wanders into a restricted zone, there's no human guard to notice and fix it immediately.

Enter SafeGuard ASF. Think of this not as a robot, but as a super-heroic, walking security guard designed specifically for these empty factories.

Here is how it works, broken down into simple concepts:

1. The Body: The "Walking Human"

Instead of using a wheeled robot (which gets stuck on stairs) or a drone (which can't open doors or climb ladders), this system uses a humanoid robot (the Unitree G1).

The Analogy: Imagine a robot that looks and moves like a person. Because it has legs and a human shape, it can walk up stairs, squeeze through narrow hallways, and reach high shelves where dangerous pipes are located. It fits perfectly into a world built for humans, without needing to remodel the factory.

2. The Eyes and Brain: The "Super-Senses"

This robot doesn't just look with regular eyes; it has a multi-sense superpower.

The Eyes (RGB-D Camera): It sees the world in 3D, like a human with depth perception. It can spot a fire or smoke instantly.
The Thermal Vision (Infrared): It can "see" heat. If a pipe is getting too hot (like a fever in a human body), the robot spots it before it bursts.
The Brain (AI Reasoning): This is the coolest part. The robot doesn't just sound an alarm like a simple smoke detector. It uses a "Reasoning" brain (called ReAct).
- The Metaphor: A smoke detector is like a dog that barks when it smells smoke. SafeGuard is like a detective. It smells the smoke, looks around to see where it is, checks a map to see what's nearby, and decides: "Oh, this fire is near a chemical tank! I need to call the hazmat team, not just the fire department." It makes smart choices, not just automatic reactions.

3. The Toolkit: The "Swiss Army Knife"

The robot has a digital toolkit with 23 different tools it can use.

It can scan for fires.
It can check if a person is authorized to be there (like a bouncer checking IDs).
It can map out the factory in its memory.
It can even talk to a central control room to say, "Hey, I found a problem, here is exactly where it is."

4. The Training: The "Robot Gym"

Before this robot ever stepped foot in a real factory, it went to a virtual gym (a computer simulation).

The Analogy: Think of it like a video game character training in a simulator. The robot practiced walking, running, and even doing a "Gangnam Style" dance (to prove it could balance perfectly) millions of times.
It learned how to walk on uneven floors, how to recover if it got pushed, and how to navigate obstacles without falling over. Once it mastered the game, they transferred those skills to the real robot.

5. What Happens When Things Go Wrong?

The paper tested the robot on three main dangers:

Fire/Smoke: The robot spots the fire, figures out how bad it is, and tells the control center to shut off power or spray water.
Overheating Pipes: It spots a pipe getting hot, tracks the heat over time, and predicts when it might explode, giving humans time to fix it.
Intruders: If someone who shouldn't be there walks in, the robot recognizes them, checks if they are an employee, and if not, alerts security.

The Bottom Line

SafeGuard ASF is like giving the factory a self-driving, walking, thinking guardian.

Old Way: You have fixed cameras and sensors that might miss things or can't move.
New Way: You have a robot that patrols the whole factory, climbs stairs, sees heat, thinks about the problem, and takes action.

The tests showed that this robot is incredibly fast (spotting fires in less than a tenth of a second) and very accurate (getting it right about 94% of the time). It proves that in the future, our factories won't just be run by machines; they will be guarded by them, keeping everything safe even when no humans are watching.

1. Problem Statement

The manufacturing industry is transitioning toward "dark factories"—fully automated facilities operating without human presence. While this improves efficiency, it creates critical safety gaps:

Limitations of Current Systems: Traditional safety relies on fixed sensors (thermal cameras, gas detectors) which suffer from coverage blind spots, high false alarm rates, and an inability to perform physical intervention.
Multi-Hazard Complexity: Industrial environments face simultaneous threats including fire, smoke, gas leaks, equipment overheating, and security breaches. Current solutions are fragmented, addressing only one hazard type at a time.
Need for Active Response: Existing automated systems often lack the mobility to navigate complex human-designed environments (stairs, narrow passages) or the reasoning capabilities to make context-aware decisions during emergencies.

2. Methodology

The authors propose SafeGuard ASF (Agentic Security Fleet), a comprehensive framework deploying the Unitree G1 humanoid robot for autonomous hazard detection and response. The system is built on a six-layer hierarchical architecture:

A. Hardware Platform

Robot: Unitree G1 Humanoid (29 degrees of freedom).
Sensors: Intel RealSense D455 (RGB-D) and thermal imaging capabilities.
Compute: NVIDIA Jetson Orin (275 TOPS) for on-board processing.

B. Locomotion and Control

Training: Utilizes Unitree RL Lab with Proximal Policy Optimization (PPO) in Isaac Sim.
Simulation-to-Real: Employs massive parallel training (4,096 environments) with domain randomization (friction, mass, motor strength, external pushes) to ensure robust transfer to real hardware.
Capabilities: Trained policies support walking, running, and complex motion tracking (e.g., dance) with stable convergence within 80,000 iterations.

C. Perception Pipeline (Layer 1 & 2)

Fire/Smoke Detection: Fine-tuned YOLOv8-m model achieving 94.2% mAP with 127ms latency. Severity is estimated based on detection area and confidence.
Thermal Anomaly Detection: Compares real-time thermal scans against stored baselines to detect pipeline overheating ( $\Delta T > 15^\circ C$ ). Includes trend analysis to predict time-to-critical-threshold.
Intruder Detection: Uses YOLOv8 for person detection and OSNet for Re-Identification (Re-ID) to distinguish authorized personnel from intruders.
3D Localization: Fuses depth data with camera intrinsics to localize hazards in 3D space.

D. Agentic Reasoning (Layer 4 - ToolOrchestra)

Framework: Adapts the ReAct (Reasoning and Acting) paradigm for real-time robotics.
ToolOrchestra: A structured decision-making system comprising 23 specialized tools across four categories:
- Perception (e.g., fire scan, thermal mapping)
- Reasoning (e.g., severity assessment, root cause analysis)
- Knowledge (e.g., facility maps, personnel DB)
- Actuation (e.g., remote valve control, suppression triggers)
Workflow: The robot cycles through Thought (analyze observation), Action (execute tool), and Observation (update state) to generate context-aware responses (e.g., prioritizing power isolation for electrical fires vs. hazmat protocols for chemical fires).

E. Planning and Action (Layers 5 & 6)

Navigation: Hierarchical planning using D Lite* for global pathfinding and MPPI (Model Predictive Path Integral) for local trajectory optimization.
Actuation: Interfaces with the robot's locomotion controller and communication systems to execute alerts, verbal warnings, and remote interventions.

3. Key Contributions

First Humanoid Framework for Dark Factories: Introduces a system specifically designed for multi-hazard monitoring in unmanned industrial settings, addressing the unique challenges of operating without human oversight.
Integrated Multi-Modal Perception: Achieves simultaneous detection of fire, smoke, thermal anomalies, and security threats with high accuracy (94.2% mAP for fire/smoke) and low latency (127ms).
ReAct-Based ToolOrchestra: Adapts the ReAct paradigm for robotics, enabling structured decision-making via 23 specialized tools, moving beyond simple rule-based triggers to intelligent, context-aware responses.
Robust Locomotion Training: Demonstrates successful sim-to-real transfer of complex locomotion policies (including dance and velocity control) on the Unitree G1, achieving stable walking on uneven surfaces and stairs.

4. Experimental Results

Detection Performance:
- Fire/Smoke: 94.2% F1-score, 127ms latency.
- Thermal Anomaly: 91.5% F1-score, 83ms latency.
- Intruder: 95.9% F1-score, 152ms latency.
Scenario Success Rate: Across 20 trials per scenario, the system achieved an 89.3% overall success rate in end-to-end hazard response (Fire: 92%, Thermal: 88%, Intruder: 88%).
Response Time: Average end-to-end response time is 12.4 seconds for fire scenarios. Navigation dominates this time (55-70%), while the ReAct reasoning phase adds ~0.7–1.2 seconds.
Case Study: Successfully detected a stuck valve causing a 60.8°C temperature spike on a pipeline, diagnosed the root cause, and triggered a remote reset, returning the system to normal within 12 minutes.
Real-World Validation: Demonstrated autonomous patrol, obstacle avoidance, human detection, and social interaction (handshake) on physical hardware, confirming successful sim-to-real transfer.

5. Significance and Impact

Paradigm Shift: Moves industrial safety from passive, fixed monitoring to active, mobile, and reasoning-driven guardianship.
Environmental Adaptability: The humanoid form factor allows the robot to navigate human-designed infrastructure (stairs, narrow aisles, elevated racks) that wheeled or quadruped robots cannot easily access, eliminating the need for facility modifications.
Scalability: The "Agentic Fleet" concept suggests a scalable approach where multiple robots can coordinate to cover large facilities, addressing the limitations of single-point sensors.
Future of Automation: Validates the viability of humanoid robots not just for manufacturing tasks, but as critical safety infrastructure in the era of "dark factories," combining the mobility of legged robots with the cognitive capabilities of agentic AI.

Limitations Noted: The current system relies on remote actuation rather than physical manipulation (e.g., turning valves manually), has limited multi-robot coordination in the current study, and faces challenges in extreme lighting or heavy smoke occlusion. Future work aims to address physical manipulation and fleet coordination.