Humanoid GR-1 robot only perceives environment via cameras
Fewer sensors also means cheaper - and better, claims Fourier Intelligence. The company relies solely on cameras for its humanoid GR-1 robot.
(Image: Fourier Intelligence (Screenshot))
The robotics company Fourier Intelligence has redesigned its humanoid robot GR-1 and simplified its perception system. The robot now manages without a large sensor package to detect its environment. Instead, Fourier relies on a camera array to enable a 360-degree all-round view - and on a neural network to evaluate and translate the camera images into usable environmental data.
Fourier calls the new perception system "Pure Vision", which enables the humanoid GR-1 robot to perceive its environment in a similar way to a human. The robotics company has deliberately dispensed with radar and lidar sensors in order to avoid making the entire technology unnecessarily complicated and to reduce costs.
Pure Vision in the revised GR-1 is based on six RGB cameras that enable a 360-degree all-round view. This allows the robot to capture its surroundings from all angles. The camera array is located in the head of the humanoid robot and enables a map to be created from a "low" bird's eye view from this elevated position. This enables the robot to perceive the environment as a whole, making it easier to track and analyze objects, for example, in order to make informed decisions.
Using a neural network to identify spatial features
With the help of a neural network, the Pure Vision system converts the complex environment of the 360-degree view into three-dimensional spatial features using transformer models. These enable the GR-1 to understand the structure of its environment and recognize areas occupied by objects, for example. To do this, the data is translated into a three-dimensional occupancy grid using an occupancy network. This enables the GR-1 to recognize which areas are passable and which are impassable and thus navigate safely and autonomously in its environment.
(Image: Fourier Intelligence)
According to Fourier, the system has passed extensive tests. The detection of moving vehicles and pedestrians is very accurate. The system also works very efficiently, making it possible to map the surroundings and track objects in real time. Similar to electric vehicle manufacturer Tesla, which has been relying exclusively on cameras and image processing systems for its Autopilot system since 2021, Fourier can thus significantly reduce manufacturing costs. At the same time, Fourier claims that the GR-1's perception of its surroundings has improved. This means that safe and efficient operation can be achieved "with human-like precision".
The current model of the GR-1 has undergone further revisions. The open, skeletal body has now given way to a closed body. With its 54 degrees of freedom throughout its body, the humanoid robot should be able to walk like a human and walk safely over various surfaces. Adaptive balance algorithms help with this.
In addition to the Pure Vision system, the previously empty head now contains loudspeakers and microphones as well as a display as a face.
(olb)
