Robbyant will advance 15 teams to the Shanghai final of its first Embodied AI Challenge, where competitors will have 48 hours to move their own models onto physical robots. The company confirmed this to Humanoids Daily, adding hardware details to its 27 October announcement of more than 1,800 registered teams from 13 countries.
Key takeaways
- The final runs 14–16 November in Shanghai as a 48-hour hackathon on physical hardware
- The platform is AgileX Robotics' Cobot Magic — a dual-arm mobile manipulation system
- The contest is built on the open LingBot-VLA 2.0 model, pre-trained on 50,000 hours of robotic data
- More than 1,800 teams registered, but the number of qualifying submissions has not been disclosed
- Final tasks, scoring criteria and the extent of advance robot access are still being settled
Picture this situation
A team fine-tunes a model for the task “pick up the mug and place it on the tray”. In simulation the success rate reaches 95 percent. The numbers look publication-ready.
On a physical robot the same model misses every other attempt. The mug reflects a lamp in a way the simulated camera never modelled, control latency is several dozen milliseconds higher, and a joint has backlash that the model did not contain.
But here is the problem: that gap shows up in no leaderboard computed in simulation.
That is exactly the point of a final where teams get 48 hours to adapt a model to a real machine. The figures in this example are illustrative — the organiser has published no results.
What does the challenge look like?
The contest is organised by Robbyant, the embodied AI company owned by Ant Group, together with ModelScope and Alibaba Cloud Tianchi. It launched in September — its announced scope covers model training and fine-tuning, generalisation in simulation, and adaptation to real robots.
The 14–16 November event will use Cobot Magic from AgileX Robotics, a dual-arm mobile manipulation platform. That hardware choice makes the final a mobile-manipulation contest rather than a comparison of humanoid bodies.
Tony Peng, PR advisor at Ant Group, wrote in responses to Humanoids Daily that finalists would need to “adapt, optimize and deploy their models on physical robots”. Qualifying submissions are still under review, so the 15 finalists have not been named.
This is a simplified view based on the announced scope — the last step stays undefined because the organiser has not yet published scoring rules.
What is LingBot-VLA 2.0?
The contest's base model is LingBot-VLA version 2.0, a vision-language-action model — it takes an image and a language instruction and returns actions for a robot.
According to the maker's technical page, pre-training for the second version used a broader data mixture: roughly 50,000 hours of real robotic data plus roughly 10,000 hours of embodiment-free egocentric manipulation footage.
That second pool is the more interesting one. First-person recordings show human hands performing tasks and need no robot hardware to collect. In exchange they have to be cleaned — the maker describes stage-by-stage filtering that rejects material with unstable camera tracking, discontinuous hand trajectories or too few frames with a visible hand.
What matters for competitors is that they start from a model already trained at that scale. The contest is therefore not about pre-training but about adaptation — precisely the part that decides deployment in practice.
Why is a simulation leaderboard not enough?
Most comparisons of robotic models today happen in simulation, because it is cheap, repeatable and can be run in parallel for hundreds of teams. The first stage of this contest works exactly that way.
The problem is that simulation does not model some of the things that actually break execution: light reflections, joint backlash, control latency, differences in friction. Moving a model onto a physical machine — sim-to-real — is a separate skill, not a formality after a good score.
So the simpler option, settling the contest on a simulation leaderboard, would measure something other than what the contest promises. A final on physical hardware shifts the weight onto adaptation.
There is a second problem the format does not solve by itself. Success after many attempts on a known task means something different from handling an unfamiliar object or a new setting. Without repeated trials and without accounting for human intervention, a successful attempt cannot be distinguished from consistent performance.
The number of finalists is also worth noting. Fifteen teams is a set small enough that results will read individually rather than as a distribution. That is good news for transparency and bad for statistical inference — at that sample size the gap between first and third place may come from task selection rather than adaptation quality.
What do these names mean?
Three terms needed to read the contest rules:
- VLA — a model taking an image and a natural-language instruction and returning actions. The equivalent of a language model, except the output is motor movement rather than text.
- Egocentric data — first-person recordings showing human hands at work. They are collected without a robot, so they scale more cheaply than teleoperation.
- Mobile manipulation — grasping combined with driving, meaning a robot that reaches the object by itself. Cobot Magic does this with two arms on a wheeled base, without legs.
That last point marks the limit of what the contest can show: it will reveal how well a model adapts to hardware, but says nothing about bipedal locomotion.
Is it working yet?
The contest is running, the final is ahead, and its scoring rules have not been settled. Robbyant said task design and evaluation criteria are still under development. It has not confirmed whether finalists will meet unfamiliar objects, environments or tasks, or how much access to robots and real-world training data they will have beforehand.
Asked about improvements over the base LingBot-VLA 2.0 and about publishing baseline comparisons, results and complete-attempt videos, Peng said evaluation was ongoing and further details would follow. That response contained no commitment to publish those materials.
The registration count signals substantial interest, but it should not be read as 1,800 completed qualifying entries.
Why it matters
Open robotic models are today compared mostly on their makers' own benchmarks. A contest in which external teams adapt the same model on identical hardware within an equal time limit offers a rare chance to compare adaptation skill itself. Its evidential value, however, depends entirely on rules that have not yet been announced.
What's next
- The final takes place 14–16 November in Shanghai on the Cobot Magic platform
- The organiser has yet to announce final tasks, scoring criteria and the extent of advance access to robots and data
- There is no commitment to publish baseline comparisons, results and complete-attempt videos — those will decide whether the results are checkable
Sources
- Humanoids Daily — Robbyant's Embodied AI Challenge Will Send 15 Teams to a Real-Robot Final in Shanghai
- Robbyant — LingBot-VLA 2.0





