Galaxy General Sweeps Scenario Golds at Humanoid Robot Games with Full Autonomy

The scenario competitions were widely seen by industry insiders as a far more rigorous test than track-and-field events. While track events only assess motion control algorithms, scenario competitions require robots to handle dynamic environments, soft objects, and interruptions. This year, the organizers allowed teleoperation, but the scoring rule gave autonomous operation a weight of 1.0 versus 0.5 for teleoperation, and ranked autonomous teams first in ties. Despite this, most teams chose teleoperation to avoid risks, as a player from a major embodied AI company told AI Tech Review: "A 0.5 discount is painful, but it is better than having the robot crash on the spot or smash the kitchen."

Galaxy General, however, insisted on full autonomy in all three events. The robots were powered by the company's self-developed AstraBrain world action model, which handled both decision-making and execution without any human intervention. In the household scenario, the robot completed the longest course of about 30 minutes early; in the retail scenario, the team swept gold, silver, and bronze. The final result was a 100% win rate across the three gold medals.

The household scenario required the robot to move between rooms, pick up objects, open doors, load a washing machine, fold clothes, and hang them. Two technical difficulties stood out: manipulating soft objects like clothes, which have infinite degrees of freedom and require real-time visual-motor adjustment, and recovering from unexpected interruptions. At one point, a random voice command told the robot to bring in a package from outside the door. The robot had to interrupt its work, complete the new task, and then resume the previous step — a challenge involving memory and time-space grounding that often leads to cascading errors. Galaxy General's robot completed all interference tests and scored a perfect score on the first technical review.

In the catering scenario, the robot had to use food tongs, operate a microwave knob, and deliver drinks. Tool use and compliant force control are classic manifestations of Moravec's paradox. The robot finished with zero mistakes in 6 minutes 55 seconds, the fastest among all teams.

The retail scenario was a 20-minute timed test involving picking goods according to online orders, patrolling shelves, replenishing stock, and correcting misplaced items. The dynamic orders and visually similar products made pre-programmed rules useless. Galaxy General not only completed the tasks but also took all three podium places.

Galaxy General's performance did not come from a competition-specific model. The company operates humanoid-run retail stores, called "Galaxy Space Capsules," in nearly 50 cities at about 200 locations, with stable operation for more than a year. It also built nearly a hundred smart pharmacies, the first opened in September 2024, which has been running 24/7 in a dense space for almost two years and has dispensed over one million drug packages. These real-world interactions generate physical feedback data that is continuously fed back into AstraBrain, creating a closed loop from real application to data to model evolution.

The results point to a broader shift in the humanoid robot industry. As hardware becomes increasingly homogeneous, competition is moving to the embodied intelligence brain. Wang He, founder and CTO of Galaxy General, has said that embodied intelligence is approaching its own "ChatGPT moment," marked not by memorized actions but by genuine generalization. The three full-autonomy gold medals at this competition provide fresh evidence that such a moment may be near.