Liquid AI runs warehouse-inspection VLM on AMD
Liquid AI, AMD and Robotec.ai showed a 3-billion-parameter vision-language model doing warehouse safety inspection entirely on-device, an example of agentic robotics running without a network connection.
Liquid AI announced a joint demonstration with AMD and Robotec.ai. In it, the company's LFM2-VL-3B vision-language model runs locally on AMD Ryzen AI processors and handles perception and decision-making for an autonomous inspection robot. The work was set for presentation at ROSCon 2025 in Singapore.
The prototype is a mobile robot in a simulated warehouse that has both human and machine traffic. All inference runs on the robot with no internet or network connection, using llama.cpp with native AMD acceleration. The model takes natural-language instructions and raw images as input. It returns structured JSON that flags package defects, hazards, OSHA violations and operational anomalies, and recommends corrective actions, all in one inference pass. In Liquid AI's example, the model spots a fallen box in an aisle, records it as a safety violation, and recommends removing it. Downstream agents can parse that output and act on it.
Robotec.ai validated the system in a ROS 2-based hardware-in-the-loop simulator connected to the robot's sensor and actuator interfaces. The team says it ran hundreds of warehouse scenarios there, testing decision accuracy, replanning and perception reliability before any physical deployment. The agent layer runs on Robotec.ai's RAI framework.
The post gives no figures for latency, throughput or accuracy. It says only that "preliminary measurements" show the stack works within the real-time limits typical of closed-loop industrial systems.
The demonstration builds on Liquid AI and AMD's earlier LEAP SDK integration. LFM2-VL-3B is available on Hugging Face under the LFM Open License and through the LEAP SDK. Liquid AI says it is providing code examples and benchmark data for research and integration.