This video unboxes and reviews the Nvidia Jetson AGX Thor developer kit, positioning it as the most powerful brain available for robotic applications. It details the kit's hardware specifications, including its Blackwell architecture, extensive CUDA and Tensor cores, 14-core ARM CPU, and a massive 128GB of memory. A comparison with the previous Jetson AGX Orin highlights the Thor's enhanced AI performance and power efficiency, attributing its larger size primarily to a sophisticated heatsink design. The video demonstrates the Thor's capabilities through three advanced AI applications: a "hardware-in-the-loop" robotic simulation using a visual-language-action (Groot) model, a video search and summarization system for safety alerts, and local execution of large language models (Llama 3 70B and 23B parameters). The presenter discusses integration benefits like single-side port placement and the potential for custom cooling solutions, alongside noting some I/O differences compared to the Orin.
Summarizing the Jetson AGX Thor Developer Kit for Robotics
0:02 Introduction to Jetson AGX Thor: The video introduces the Nvidia Jetson AGX Thor developer kit as the most powerful embedded AI platform for robots, suitable for humanoid, mobile, and self-driving car applications. It boasts a Blackwell architecture, 2560 CUDA cores, 96 Tensor cores, a 14-core ARM CPU, and an impressive 128 GB of memory.
0:48 Hardware Overview & Unboxing: The kit features a substantial heatsink with a fan, an SSD with its own heatsink, and a Wi-Fi/Bluetooth card. Key ports include DC/USB-C for power, HDMI, DisplayPort, Ethernet, USB, and a high-speed QSFP28 port. All main ports are strategically located on one side for simplified integration into robotic systems.
1:39 Comparison with Jetson Orin & Design Philosophy: When compared to the Jetson AGX Orin, the Thor developer kit is physically larger, primarily due to its more advanced heatsink with heat pipes. However, the core Jetson module itself maintains a similar form factor to the Orin. The video suggests removing the heatsink and mounting the Jetson directly to a robot's chassis for cooling, reducing the overall weight and size for embedded applications.
4:12 Hardware-in-the-Loop (HIL) Demo Concept: The presenter explains the "Hardware-in-the-Loop" methodology, where a robot's perception and policy (AI inference) run on the real Jetson hardware, while the robot's physical environment and sensor data are simulated on a powerful PC. This setup allows for realistic testing of AI models without a fully built physical robot.
6:32 Groot Model HIL Demonstration: A practical HIL demo is conducted using the Groot visual-language-action model. The Jetson Thor handles the AI inference to control a simulated robot pouring nuts between containers and placing them on a tray, with randomized object positions to test robustness. ROS (Robot Operating System) facilitates communication between the Jetson and the workstation running the simulation.
12:18 Video Search & Summarization: Another demonstration showcases the Thor's capability in video analytics for safety. It successfully identifies a person on a ladder without personal protective equipment (PPE) in a video clip and provides context-aware answers to safety-related questions.
14:27 Local Large Language Model (LLM) Execution: The Jetson Thor is shown running large language models locally, specifically Llama 3 (70 billion and 23 billion parameters). While the 70B model's response is slow, it is accurate, highlighting the Thor's power for demanding on-device AI workloads. The 23B model provides a faster, albeit less detailed, response.
16:00 Key Takeaways & Future Plans: The Jetson AGX Thor offers a significant performance leap (7.5x more AI performance) and improved efficiency (3.5x more efficient) compared to the AGX Orin. Despite its larger heatsink, the core module size is similar. Noted disadvantages include the absence of additional M.2 slots, a dedicated camera connector, and the GPIO 40-pin connector found on the Orin. However, the QSFP28 port provides extremely high-speed communication potential for motors and sensors. The presenter plans to utilize the Jetson Thor for future robotics projects, including a two-arm robot torso and legged robots.