Advancing Physical AI with AMD and Liquid AI
Goals:
- Build a high-fidelity warehouse simulation for navigation, perception, and manipulation benchmarks
- Integrate and optimize Agentic AI (RAI + VLMs) with ROS, MoveIt, and robotic control stacks
- Achieve embodied autonomy capable of task sequencing, inspection, manipulation, and reporting
- Deploy real-time, on-device AI inference on AMD Ryzen AI hardware
Case study: Advancing Physical AI with AMD and Liquid AI
Partners: AMD, Liquid AI
Challenge
AMD asked us to develop a novel Edge AI Hardware-in-the-Loop (HiL) simulation with Agentic AI to perform autonomous robotic tasks in a warehouse using a mobile manipulator (Robotnik). Liquid AI engaged with us to apply and fine-tune their Edge AI models for robotics and specific domain. The challenge focuses on achieving Embodied AI combining perception, reasoning, and manipulation while ensuring hardware-efficient real-time execution on AMD Ryzen AI platforms.
- High-fidelity simulation: Build a dynamic warehouse environment with diverse assets to benchmark AI-driven, flexible navigation, perception, and manipulation, and provide a fine-tuning platform for VLM.
- Agentic AI integration: Apply and optimize the agentic AI framework RAI to combine VLMs with robotic control stacks (ROS, MoveIt), enabling adaptive task planning and introspection.
- Embodied intelligence: Implement fully on-board, closed-loop autonomy capable of task sequencing, navigation, inspection, manipulation, and reporting driven by multimodal reasoning.
Solution
- High-fidelity simulation & benchmarks: A simulated warehouse world with safety and compliance violations (e. g. spills, blocked exits, violations).
- Agentic AI stack (RAI + Liquid LFM2-VL): Robotec applies the RAI multi-agent framework with Liquid AI’s LFM2-VL and open-source LLM to couple perception and reasoning with ROS 2 packages (MoveIt, Nav). The agent toolbox covers navigation, vision, manipulation, reporting, and state introspection.
- Hardware-in-the-Loop (edge deployment): The full stack is deployed on a Mini-PC with AMD Ryzen AI, running ROS nodes, ONNX-runtime models, and targeted RAI for real-time inference and diagnostics.
Results
- The first successful case of multi-agent embodiment fully on-board, proving huge potential of Agentic AI for robotics.
- Huge VLM performance gain with synthetic data fine-tuning We delivered high-quality datasets that improved VLM performance in challenging inspection tasks drastically to 95%.