Another Chinese robotics company has just stepped into the spotlight with an unusual demonstration: placing a humanoid robot inside a go-kart and teaching the machine not only to drive but also to drift autonomously. The company is called Symbiosis Robotics, and according to released information, it was founded just about a month prior to the presentation. For the demonstration, the company used a Unitree G1, a commercial humanoid standing approximately 1.5 meters tall and weighing close to 35 kg.

In the video, the robot must adapt its body to the kart's cramped space, grip the steering wheel, reach the pedals, and continuously coordinate its vision and movements to control the vehicle. According to Symbiosis Robotics, the operation is not performed via remote control. Powering the demonstration is an artificial intelligence model known as Direct Perception Control (TPC). The aim is to reduce the traditional separation between perception and movement control.

In many robotic systems, distinct modules are responsible for perceiving the environment, planning an action, and ultimately translating that decision into joint movements. TPC employs a different approach. The model receives multimodal information—including vision, language, the robot's own body state, and tactile data—and maps it directly to the movements of the robot's joints and hands. The company describes this architecture as a way to bring the "brain" (responsible for perception) closer to the "cerebellum" (responsible for motor control).

The goal is to enable a single model to learn skills across various body types and situations; driving a go-kart offers a way to test several of these capabilities simultaneously. The humanoid must visually track the course while turning the steering wheel with its hands and operating the accelerator and brake with its feet; movements must also be continuously adjusted during a drift. Furthermore, the robot must maintain control of its own body while experiencing the forces of acceleration, braking, and changes in direction.

For Symbiosis Robotics, integrated perception-and-control models could pave the way for developing humanoids capable of learning increasingly complex physical tasks, without relying on separate architectures for each stage of the operation.