The third-place team at the first Ant LingBot Embodied AI Model Challenge collected only 13 real-robot data points and scored 65.87 points; the champion team exceeded 82 points. We believe this competition matters less for whether robots “know how to do the task” than for what it reveals: the center of gravity in embodied AI is shifting from simply scaling up data to engineering task decomposition in real environments.
What This Is
This was a 48-hour challenge conducted on physical robots. Of 1,855 registered teams, 15 reached the finals. All used the same open-source foundation model, LingBot-VLA 2.0; the same bimanual robot, COBOT MAGIC; and identical computing power to complete three tasks: placing fruits and vegetables, organizing cosmetics, and bimanual floor cleaning.
SST divided the work among different models: DeepSeek interpreted instructions, a vision model located targets, and LingBot-VLA 2.0 generated actions. The champion, TARS, first excluded irrelevant objects, then adjusted its control trajectories to compensate for the robotic arm sagging under load. Another team converted each real-robot failure into new training data, progressively raising its success rate from roughly 10%.
More critically, the simulation environment differed from the real tabletop by about five centimeters, causing robotic arms to grasp empty space or even jam. Reflections from transparent containers, friction, and robotic-arm damping were also difficult to reproduce faithfully in virtual environments.
Industry View
We believe open-source foundation models already provide basic visual and action capabilities. Yet current product differentiation comes more from task decomposition, targeted data, and on-site debugging than from having a model with more parameters. Reducing the time required to train a single step from roughly 15 seconds to 3.4 seconds may sound like a modest efficiency gain, but it allows teams to complete several more “train–deploy–debug” cycles within 48 hours.
The risks are equally clear: small-sample fine-tuning can damage general capabilities, and performance may deteriorate when the angle or lighting changes. Transparent, reflective objects, precise force control, and bimanual coordination still lack general-purpose solutions. Ant LingBot CEO Zhu Xing acknowledged that the industry has yet to identify which category of data, when scaled up, can produce a leap in capability. In other words, embodied AI has not yet developed a Scaling Law comparable to that of language models—a stable relationship between model scale and performance improvement. A single high score cannot prove that general capabilities have matured.
Impact on Regular People
For enterprise IT: When piloting robots, enterprises should first determine whether workflows can be decomposed reliably and whether sensors and operating conditions are controllable. Costs should be calculated according to the real-robot debugging cycle rather than relying solely on simulation demos or model leaderboards.
For individual careers: In the short term, the jobs most likely to grow are robot operation, scenario testing, task orchestration, and data maintenance. Repetitive tasks such as basic material handling will be affected first, while the pace at which complex jobs are replaced remains constrained by reliability and cost.
For the consumer market: In the near term, home robots will continue to focus on clearly bounded tasks such as tidying and cleaning. Beyond price, the decisive purchase factors will be failure rate, noise, maintenance frequency, and the ability to adapt to different home environments.