Ant Lingbo Holds 48-Hour Real-Robot Embodied Large Model Competition; 15 Teams Complete Three Tasks Using LingBot-VLA 2.0
The embodied large model challenge initiated by Ant Lingbo attracted 1,855 teams and 3,252 contestants to register, with 15 teams ultimately advancing to the 48-hour real-robot finals, using the open-source model LingBot-VLA 2.0 and the AgileX COBOT MAGIC robot to complete three tasks: fruit grasping, cosmetics storage, and dual-arm sweeping. Contestants were generally constrained by the insufficient coverage of the official 647 public data entries, needing to collect additional data via teleoperation themselves, and to address real-robot issues such as transparent box recognition, lighting changes, and robotic arm sagging. The champion team TARS used Qwen to identify targets and distractors, established load deviation compensation, and reduced training time per step from about 15 seconds to 3.4 seconds. Ant Lingbo also released the LingBot developer ecosystem plan, initially inviting 20 MVPs and recruiting 40 developer experts.
Why read this
AI score 54
Through the specific process of the 48-hour real-robot finals with 15 teams, this presents the real deployment challenges and solutions encountered by different developers using the same open-source embodied model.
Topics
Body · AI translation
Beyond the Leaderboard: A 48-Hour Real-Robot Contest Tests the True Mettle of Embodied Large Models
Zhidongxi Download the App | Exclusive First Look November 16, 2026, 19:58:21 | From Beijing
Since the start of this year, Benchmark leaderboards in the embodied intelligence industry have been especially lively: new models are released one after another, and SOTA keeps being refreshed.
Take the real-robot evaluation platform RoboChallenge as an example. In the year since it launched, the top spot on Table30 has changed hands several times, and the highest task success rate has climbed from just over 40% to nearly 70%. In various robot demo videos, the motions are also becoming increasingly fluid.
The value of a leaderboard lies in comparability: under unified tasks and conditions, it measures progress in model capability. But the industry also cares about another question:
If a model is handed to external developers—a different group of people, a different batch of data, a different operating environment—what would its success rate look like then?
Today, a 48-hour real-robot final in Shanghai put this question to developers.
This embodied large model challenge, initiated by Ant Lingbo, attracted 1,855 teams and 3,252 contestants from 13 countries. After an online simulation preliminary round, 15 teams ultimately advanced to the real-robot final, where they jointly used the open-source embodied large model LingBot-VLA 2.0 and the Songling COBOT MAGIC robot to complete model tuning, real-robot deployment, and task challenges within 48 hours.
A leaderboard can reflect a model's level under given conditions, but how much data developers can use and how much time they can spend to turn those capabilities into stable real-world operation is closer to the industrial value of embodied large models.
1. From simulation to real robots: what happened in 48 hours?
At the final venue, we found that the participants' backgrounds were quite diverse: there were embodied intelligence researchers, cross-disciplinary engineers from internet algorithms and autonomous driving, undergraduates and big-tech employees touching a real robot for the first time, and young developers who formed teams on the spot via Xiaohongshu...
The open-source base model gave these developers a common starting point for cross-disciplinary exploration and also provided new opportunities to validate their existing technical experience.
For example, team "fbfyfbf" built a training environment with the help of open-source documentation and an AI coding assistant; team "416 Ding Neng Xing" introduced methods from the search and recommendation field into robot development; and team "You Shou Jiu Xing" tried to transfer technical experience from the autonomous driving field. Based on LingBot-VLA 2.0, they were able to quickly enter the embodied R&D phase.
But once they truly stepped onto the real-robot arena, the difficulties quickly became concrete. The final had three tasks: first, grasping fruits and vegetables with a single arm; second, grasping cosmetics with a single arm and placing them into a transparent box; and third, using both arms together to sweep desktop trash and then pour it into a storage basket. These tasks posed considerable challenges to the robot's visual understanding, spatial perception, and motion execution.
The first hurdle was data. Some contestants mentioned during the roadshow that the organizers provided 647 public data entries, but these did not cover all task situations. For example, in fruit placement, objects such as star fruit and green chili peppers appeared without corresponding demonstrations, so contestants could only collect additional data themselves via teleoperation.
However, more data is not always better; what to collect and how to collect it are equally critical. Team "416 Ding Neng Xing" added more than 100 data entries during the final, but because the collected content did not match the evaluation tasks well enough, their subsequent results actually failed to surpass the first round. Other teams also found that collecting additional data for failed actions, difficult objects, and weak links in tasks was often more effective than blindly expanding the data scale.
The second hurdle came from the real physical environment. Many problems were not obvious during the simulation stage, and quite a few teams were touching a real robot for the first time, so these challenges exceeded previous expectations.
Some teams reported that the spatial layers of the transparent storage box were difficult to identify accurately, and on-site lighting changes could also affect the robot's visual judgment; there were also robots repeatedly failing to grasp fruits, the robotic arm sagging after extension, and the sweeping task being unable to smoothly transition to the pouring action after the first half was completed... A tiny environmental or motion deviation could affect the execution of the entire task set.
Even trickier, all problems had to be solved within 48 hours. The Chixu Zhixing team said that in order to put the hundreds of data entries collected on site into training more quickly, the team optimized the training process and compressed the time per step from about 15 seconds to about 4 seconds.
With tight time, limited data, and an on-site environment full of uncertainty, the development pressure shown in this competition is quite similar to the constraints faced in real industrial deployment. How to use limited resources to find the most effective direction for improvement as quickly as possible became a test faced by all teams.
2. The same robot brain, developed into different solutions
In this high-pressure test, the top three teams offered some highly representative solutions centered on data feedback loops, visual prompts, and task-specific optimization.
During testing on the first day of the competition, Team EVA found that the initial model trained only on official public data had a success rate of only about 10% on the final tasks. The team judged that one of the key factors affecting model performance was the gap between the training data and the actual evaluation scenarios. Therefore, instead of relying solely on adjusting model parameters, they built an RSI mechanism capable of continuously improving data quality.
Within 48 hours, Team EVA completed about seven rounds of data feedback, added more than 700 data entries, and at the same time enhanced model capabilities using object localization, segmentation annotation, and depth information distillation.
Team SST chose to let the robot first see clearly what it needs to operate on. Drawing on the division-of-labor idea of System 1 and System 2, they used an external model to parse task instructions, then used SAM 3 to localize target objects and generate visual masks, weakening background interference so that LingBot-VLA 2.0 could focus its attention on the objects that actually need to be grasped or placed.
Cleverly, Team SST also generated segmentation data for target objects by reading videos in reverse, using it to optimize model training and action execution. At inference time, the model generates 50 actions at once but executes only 16 of them, to improve the smoothness and precision of robot operation.
The champion Team TARS, by contrast, split the three tasks and handled them separately, concentrating resources on different difficulties. In the fruit placement and cosmetics storage tasks, the team used Qwen to identify targets and distractors, weakening irrelevant backgrounds in the images to reduce the interference of cluttered environments on model judgment.
Team TARS also found that in the garbage cleaning task, the further the robotic arm extends forward, the more prone it is to sagging under load. The team tested by adding loads, established a deviation compensation method for different extension states, and corrected actions during the inference execution stage. At the same time, the team shortened the per-step training time from about 15 seconds to 3.4 seconds, improving model iteration speed.
The three teams took different paths, but all built on the same open-source foundation model, each working on data, perception, and execution. For developers, this means they can spend their limited time on the difficulties of specific tasks, and the problems they encountered and the solutions they provided also offer references for subsequent model improvements.
III. Making models truly usable: embodied intelligence needs its own developer ecosystem
The 48-hour competition is over. How can these technical attempts and practical experiences continue to play a role? How can more developers continue to explore on the foundation laid by those who came before?
"We care both about how many people participate and about what remains after participation," Ant Lingbo said, expressing the hope that the model usage experience, technical methods, and engineering results accumulated during the competition can continue to enter teaching, research, and actual projects, becoming resources that other developers can reuse.
To this end, Ant Lingbo released the LingBot developer ecosystem plan. The first phase plans to invite 20 LingBot MVPs (Most Valuable Professionals) and recruit 40 LingBot Developer Experts.
The former mainly targets university professors and researchers, promoting the entry of the model into course construction, academic research, and talent cultivation; the latter mainly targets core developers and technical creators, encouraging the development of demos, open-source engineering, evaluation scripts, and the sharing of tutorials and practical experience.
According to the introduction, Ant Lingbo will provide funding support for the plan, as well as opportunities for exchanges with the core R&D team, project cooperation, and joint laboratory construction.
Developers' practice will in turn drive the development of foundation models. Ant Lingbo believes that "good feedback is not necessarily praise; it can also be a repeatedly verified problem." By applying LingBot to different tasks and usage environments, developers can discover problems the model has not yet solved, and this feedback can help the R&D team clarify improvement directions and continuously refine model capabilities and development tools.
This is also the developer ecosystem flywheel Ant Lingbo hopes to build: developers create applications and accumulate experience based on open-source models, and the relevant results are used by more people through courses, tutorials, open-source projects, and other means; new practice in turn brings new problems and needs, driving further model iteration and attracting more developers to participate.
Real-world tasks and scenarios are extremely complex, and it is difficult for any single model R&D team to cover all application needs. Only by enabling developers in different fields to continue innovating around open-source foundation models and precipitating successful experience into reusable resources can robot brains have the chance to enter more real-world scenarios.
Conclusion: The "Android moment" of embodied intelligence is brewing
In the industry, "Android moment" is often used to describe the ecosystem effects brought by an open foundation platform. That moment clearly cannot be declared to have arrived on the strength of a single model open-sourcing or a single developer competition. But this attempt by Ant Lingbo has already shown the outside world this possibility.
48 hours, 15 teams, three tasks—the sample is not large, yet for the first time it put the question "how far can the same model go in different people's hands" on the table.
This is precisely the meaning of an open-source foundation model. It cannot guarantee that robots will perform excellently in any scenario, but it gives more people the opportunity to personally discover where the model's boundaries lie and how to push those boundaries outward bit by bit. Every recorded failure case and every shared supplementary data collection method is a path later arrivals do not have to walk again.
If more and more developers in the future can innovate based on open-source embodied foundation models and give experience back to the community, innovation in embodied intelligence will no longer occur only in a few laboratories.
The power to change the industry will also spread to countless developers.
"Special statement: The above content (including videos, pictures, or audio if any) is uploaded and published by users of "Dafeng Hao," a self-media platform under Phoenix Net. This platform only provides information storage space services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."