The Beijing startup robotics embodiment, X Square Robot, made available on an open origin framework called HOST, which teaches humanoid robot to execute the task on the basis of a short video of a human demonstration. According to the X Square Robot in the case of HOST-using robot learning via a 29-second video clip, the device achieves 62 percent performance of the task in the first approach. For comparison, the standard method in the manufacture requires about 50 expert demonstrations and 4 hours of model learning, yet it only gives 38 percent of the effectiveness.
To clarify: the current standard is based on telesurgery. An engineer with VR goggles and equipped with wire or wireless manipulators to execute a circumstantial task (for example, “put the flower into a vase”) leads the robot arm step by step, repeats series dozens of times, after which the model is taught for at least respective hours by supervised learning. A single “tudent” robot to collect data costs hundreds of thousands of dollars, and the marginal cost of 1 demonstration could exceed a 100 dollars. Indian “data factories” and casual workers from Chinese region towns are presently the cheapest origin of traffic data. However, given the number of repetitions and variants needed, the low-cost origin of traffic data does not solve the problem of scale. The learning process is presently expensive, tedious and does not bring spectacular results.
HOST reverses the logic of the most popular humanoid manufacturers approach. alternatively of following human movements, the robot “imaginates” the final state of the task and calculates the essential movement himself. The video module reproduces the scene from a robot perspective, the action module calculates the trajectory of arms, forearms, palms and fingers, and the frame matching mechanics by “dynamic time warping” reduces the adjustment mistake almost 10 times. The approach refers to cognitive simulation theory: the robot is not taught again, but treats the video as a clue at the time of the inference.
X Square Robot was established in late 2023 and has already acquired over a billion yuan in 8 rounds of funding. She was 1 of the exhibitors at the planet Robot Conference 2026 in Beijing.
Leszek B. Glass
Email: [email protected]
