Consumer‑grade Embodied AI Robot Achieves 1000× Compute, Beats Nvidia Jetson Thor for 1/10 Cost

VeilBlue has just released a new consumer‑grade quadruped robot that looks like a toy but offers server‑level performance. Compared with the previous generation, its compute capability has increased by 1,000×, matching the 2,070 TFLOPS of Nvidia's Jetson AGX Thor T5000 chip while costing roughly one‑tenth of Nvidia's price.

Heterogeneous edge computing architecture : Instead of a single‑chip solution, the robot uses a custom embodied‑intelligence edge cluster composed of six distributed chips – two 5 nm, two 8 nm, and two 3D‑stacked chips – providing a total of 22 CPU cores. Tasks are assigned to the chip best suited for them: 5 nm chips handle large‑model inference, 8 nm chips process perception data, and the 3D‑stacked chips manage motion control, with millisecond‑level task‑switching via a high‑speed bus.

Performance breakthrough : The new architecture raises the token‑per‑second (TPS) throughput for a 1.5 billion‑parameter model from the industry‑high 23 to 617, and a 7 billion‑parameter model can reach 280 TPS, making on‑device large models viable for consumer products for the first time.

"The next big thing will start out looking like a toy." – Clayton M. Christensen

Perception superiority : The robot features a 66‑megapixel, 140 dB HDR vision system operating at up to 480 fps, providing 360° coverage that exceeds human vision in resolution, dynamic range, and frame rate. Its 3D TOF and structured‑light sensors generate 2.232 million points per second, enabling centimeter‑level 3D mapping, while a 12‑mic microphone array offers precise 3D sound localization and beamforming, a first in consumer‑grade mass production.

Safety by design : Safety is integrated rather than added on. Physical safety is achieved through hidden joints and an independent safety‑brake system with millisecond response. System security combines on‑device 360° safety AI with cloud‑level protections (Alibaba Cloud, Azure). Data and privacy are protected by keeping compute local and encrypting all communications.

Autonomous operation : Unlike many robots that rely on external tags or remote commands, this model performs full‑stack perception, mapping, localization, and path planning on‑device, enabling true "see‑think‑act" autonomy. It can navigate stairs, thresholds, and slopes without human intervention, and its follow‑mode adapts to the user's pace, predicting turns and adjusting speed.

Market impact : Since the first consumer robot in 2023, VeilBlue has sold over 25,000 units, with more than 1.5 billion hours of operation, 9.5 billion minutes of usage, and 65 million interactions recorded. The company emphasizes a consumer‑first strategy, pricing the robot at a few thousand dollars to achieve scale and rapid iteration based on real‑world data.

Overall, the robot demonstrates that the critical point for consumer‑grade embodied intelligence lies not in labs but in everyday living rooms, combining high‑performance heterogeneous compute, human‑level perception, and safety‑centric design to deliver a truly autonomous household companion.