Beijing Humanoid Games Put Practical Work to the Test; Only 3 of 12 Finish Fire Course

The harder question at Beijing’s World Humanoid Robot Games was not how fast a robot could run, but whether it could complete a changing job. In a 30-minute firefighting and rescue course, only three of 12 teams finished the full sequence, according to the Global Times.

Held in Beijing from August 22 to 26, the second games mixed athletic contests with household, hotel, library, firefighting and hand-dexterity tasks. The fire course required robots to identify hazardous substances, shut three open valves of different types, find a fire, locate an extinguisher and put the fire out.

Human teleoperation was allowed in some events, with score penalties for remotely controlled robots. Teleoperation means an operator directly controls the robot’s movements, so the rules made the competition a test of both task completion and the degree of assistance behind it.

The South China Morning Post reported that some competitors described as more autonomous used 5G modules to send data to cloud-based embodied-intelligence systems and receive instructions in return. That arrangement puts at least part of the system that determines a robot’s next action outside the machine itself.

One household scenario included washing laundry, folding clothes and cleaning a living room. An unscheduled delivery then required robots to switch tasks and bring the package inside, adding a changing assignment to the set of chores.

Other work trials targeted specific capabilities

• Hotel challenges timed robots moving wheeled luggage to assigned rooms, cleaning beds and restocking supplies.
• A library event asked robots to collect returned books and place them correctly, testing visual recognition and reasoning, according to the Global Times.
• Dexterity contests included picking up beans, driving screws and hammering nails into corkboard without placing them at an angle.

X-Humanoid won the small-group 400-meter race with an unusual arms-up running posture. Its developers said the movement emerged from reinforcement-learning trials in simulation, illustrating how a robot can be trained for a defined competitive routine.

Purdue researcher Dipam Patel said robots need extensive trial-and-error training in simulation or real settings, and that covering every possible edge case across many tasks is not feasible. The important test is whether humanoids can repeatedly complete interrupted, multi-step work with less direct human control and less dependence on offboard instructions.