RaiderChip Unites Generative AI and Autonomous Robot Control on a single NPU

RaiderChip Unites Generative AI and Autonomous Robot Control on a single NPU

The new demonstrator enables intent-based voice control of Unitree robots in a virtual replica environment, combining robot policies and Generative AI models running concurrently on its NPU.

Spain, September 7th, 2026

MuJoCo is a physics simulator maintained by Google DeepMind, providing an environment for developing, training and testing robotic control systems. Robot manufacturers such as Unitree Robotics use MuJoCo to train and test policies (the AI models that control robot movement through sensors and actuators) for deployment on physical robots.

Typically, a user controls the robot within MuJoCo using velocity sliders or a joystick: the user decides what to do, the robot policy simply performs movements without falling.

RaiderChip adds a cognitive layer powered by Generative AI to recognize speech, understand user intent, and plan the movements needed to achieve a goal. These models run concurrently with the robot’s AI control policies, locally and in real time on RaiderChip’s Hardware NPU, interfaced to the demo.

Each model running on the NPU gives the robot a different capability, much like the specialized functions of the human brain. The robot AI policy acts as the motor system, executing movements and maintaining control of the body; OpenAI’s Whisper ASR model gives the ability to hear and recognize spoken language; and Qwen-3, a four-billion-parameter language and reasoning model, provides the cognitive layer for understanding the user’s high level intent, decomposing it into individual motion goals with associated path geometries, reasoning about the required low-level movements and organizing them into a schedule.

The difference becomes clear with a simple goal: “bring me that ball.” The robot policy alone cannot understand what that goal means or determine how to achieve it. In a conventional system, the user performs the planning: they must use the joystick to move the robot forward, correct its trajectory, turn it, approach the target and execute each required movement in sequence.

With RaiderChip’s Generative AI architecture, the user only needs to express the goal in natural language. The Whisper ASR model transcribes the instruction “bring me the ball”; the Qwen-3 LLM understands the goal and reasons on how to achieve it, breaking it down into an ordered, scheduled sequence of movements; and the robot policy controls the robot’s actuators, while monitoring balance through its sensors, to execute that sequence.

This demonstration highlights two fundamental aspects of the next generation of robotics:

The first is the importance of numerical precision when AI moves from the digital domain to controlling physical systems. Robot policies are trained in FP32, and aggressive quantization introduces errors that degrade motion accuracy and affect robot stability. RaiderChip’s NPU can run all models directly in their native formats, without requiring any conversion or quantization.

The second, and more important aspect is intelligence does not reside in a single model, but in the concurrent and coordinated execution of multiple specialized models. Listening, understanding, reasoning, planning and moving are distinct capabilities, enabled by Generative AI models that, working together, allow a machine to translate human intent into physical action. RaiderChip’s NPU delivers these autonomous robotic capabilities concurrently, locally and in real time, without relying on cloud-based AI services or an Internet connection.

This is where capabilities become intelligence, and intelligence becomes autonomy.

WANT TO KNOW MORE?

Learn more about RaiderChip’s NPU

Unitree is a trademark of Hangzhou Yushu Technology Co., Ltd. RaiderChip is not affiliated with, sponsored by, endorsed by, or otherwise associated with Unitree Robotics. References to Unitree products are solely for identification and demonstration purposes.