Everyone Can Build a Robot: Open Source Embodied AI With Seeed Studio | NVIDIA AI Podcast Ep. 300

Seeed Studio has supported open source hardware communities since 2008 and became an NVIDIA elite partner after seven years of collaboration that began with the Jetson Nano. The partnership provides deep technical support and faster time to market, and Seeed actively integrates NVIDIA technologies including Isaac Sim, GROOT SDK, and Jetson AI Lab into its devices as solutions, tutorials, and demos.

Eric Pan argues that open source robotics lowers barriers by letting people from different industries adopt technology without deep prior expertise, and that users trust open systems more because they can control and modify them rather than being locked into closed platforms. Seeed's customer base of millions spans hobbyists, students, researchers, and small and medium businesses, with SMBs representing a significant share of revenue from a segment that previously lacked the resources to adopt robotics.

The SO-ARM is Seeed's best-selling robotics product, developed with Hugging Face as an open source arm priced at $200. Previously, programming a robot arm required months of training to understand spatial planning and movement control. The new paradigm involves physically guiding a robot through operations several times, sending that data to the cloud for training, and deploying the resulting model on a Jetson device. A JSON file then uses a diffusion model with cameras to determine how to execute tasks, and if the result is incorrect the robot is retrained. Pan frames this as teaching a robot the way one would train a dog, positioning robots as apprentices that enhance a skilled person's work rather than replacing them and preserving the owner's proprietary knowledge.

Seeed's newer arm, Raybot, is an open source project that had reached 1,300 GitHub stars at the time of recording and is priced under $1,000. Raybot is described as more robust than the SO-ARM101 and demonstrates smooth and stable trajectory planning, with positions and actuator movements mirrored in Isaac Sim for simulation. Paired with a Jetson Nano, Raybot can run local AI functions without requiring cloud tokens. Seeed will release all 3D-printable files for Raybot parts so users can customize it for their own scenarios. Unlike the SO-ARM, Raybot is designed for applied use cases such as small business or cooking assistance, not only for learning or experimentation.

Open Claw was installed locally on a Jetson device and used to call a local API of the QwQ 3.5 model, enabling plain text commands such as move up or pick up to directly control a robot arm without coding. Approximately two weeks before the recording, Seeed connected a robot arm to Open Claw and instructed it to find its own libraries, read instructions, and build itself into the physical world. Pan describes this as giving Open Claw a physical body, allowing it not only to understand the world but to move physical objects. Control can be issued through WhatsApp messages or a microphone. A robot's behavior is defined by writing its role, definitions, and skills, making it controllable and scenario-specific, and robots running Open Claw can act as agents that collaborate with each other and support sub-agents.

Elaine Wu noted that Seeed worked with Hugging Face on the Ritchie Mini robot, going from design to manufacture in five months and shipping 3,000 units to customers. The Hugging Face LeRobot framework provides datasets, models, and policy in one end-to-end training framework without requiring users to code each component separately, and Seeed supplies compatible hardware along with customer and manufacturer services to help startups get started quickly.

Pan stated that Seeed does not want to build one general robot that does everything, but instead disassembles the humanoid concept into parts such as head, torso, arms, and wheels that people can combine according to their specific use cases. He predicted that a large fleet of physical AI creations will emerge very fast in the next one to three years. On safety, he noted that the current approach applies basic robot guidelines and uses a panic button to shut off the robot, with deeper cloud and NVIDIA integration still being explored.