The full stack behind the robot.

Embodied AI is growing fast, and every robot built for it faces the same architectural gap: the distance between raw sensors at the bottom of the stack and AI computation at the top. YUAN closes that gap with a full-stack hardware platform and the NexVDO SDK, giving robotics builders a single, integrated path from perception to decision.

Six robotic forms, one platform

The platform serves the full range of machine bodies now in development. Industrial robots need precision control on complex factory lines; humanoids demand multi-modal perception, physical coordination, and natural human-robot interaction. Quadrupeds must adapt to terrain, avoid obstacles, and run autonomous safeguarding patrols, while UAVs handle high-resolution inspection and 3D path planning. AMRs navigate warehouses and build smart maps; service robots greet, deliver, and converse. Each form draws on the same stack.

Inside the NexVDO SDK

NexVDO organises the robot's media and intelligence pipeline into five capabilities. Capture ingests data and video at low latency, with an RDMA pipeline, PTP-based sensor synchronisation, DRM, and alpha blending. Record encodes H.26x and AV1 with AES-secured storage and metadata packaging, so mission-critical inspection footage survives failure. Stream carries video over ONVIF, RTSP, and WebRTC for low-latency teleoperation. Analyze runs edge inference across vision models, speech (ASR and TTS), and modern LLMs and VLMs. Automate closes the loop with navigation, mission planning, and localisation — natively integrated with ROS2, MAVLink, visual localisation, and SLAM with NAV2.

Sensors to compute, without the wait

On the hardware side, the platform natively integrates cameras, depth cameras, LiDAR, radar, IMU, GNSS, and RTK, along with industrial protocols such as EtherCAT and motor control. The Holoscan Sensor Bridge captures high-bandwidth streams from the robot's perception modules — 10G/4K, 10G/HD — and feeds them through a Functional Safety Island directly to the main compute node, such as an Arc AGX powered by NVIDIA Jetson. The result is streaming without added latency, and a perception path the rest of the robot can trust.