Tsinghua's LoopWAM Tops RoboChallenge Table30 and Goes Open Source

The world action model LoopWAM has taken first place on the RoboChallenge Table30 leaderboard. It was developed by Tsinghua University, 原力灵机 (Dexmal) and 极寰时代. Along with the result, the authors published a project page, an arXiv paper (2610.18873) and a GitHub repository. The original Jiqizhixin report quotes only a fragment of the announcement, so it gives no exact scores and no margin over competitors. Below is what is known, and what is worth checking against primary sources.

What we know about the release

RoboChallenge was introduced as an open platform for comparing robotics models on real robot arms. Table30 is a set of 30 tabletop tasks. It is not a simulation: evaluation runs on physical robots, so first place here carries more weight than a spot on a leaderboard of virtual environments.

Key facts from the announcement:

• model: LoopWAM, of the World Action Model class;
• developers: Tsinghua University, 原力灵机, 极寰时代;
• result: first place on RoboChallenge Table30;
• license and availability: the code is open, and a paper and project page are available;
• stated philosophy: close to the concept of self-evolution during deployment, which pi is developing.

Why robots need a world model

Conventional VLA (vision-language-action) models translate an image and a text command directly into actions. World action models add an intermediate layer: they learn to predict how the scene will change after a movement. The robot does not simply react to a frame; it keeps an internal picture of the consequences.

This matters for tabletop manipulation. Pouring, folding and repositioning a fragile object are all tasks where the cost of an error accumulates step by step. If a model can forecast future states, it is less likely to end up in an unrecoverable position. The name LoopWAM hints at a closed loop between prediction and action, but the available fragment does not reveal architecture details. For those, you will need to read the paper.

The link to pi's idea

The comparison with pi is not accidental. Physical Intelligence is betting that a robot should keep learning after launch: accumulate experience, receive evaluations of outcomes and update its policy. Deployment stops being the final point and becomes a source of data.

For the Chinese team, such a parallel is convenient both in substance and for marketing. However, "does not contradict the idea" and "implements the same mechanism" are different claims. It remains to be seen whether LoopWAM learns from its own experience in real time, or whether the principle is only built into the training loop. The paper and code should answer that.

What to check in practice

For those working in robotics, it makes sense to go through a short checklist:

1. Open the project page and find the results table for each of the 30 tasks. A top overall score can hide failures on individual scenarios. 2. Compare the evaluation conditions: whether the hardware was the same, how many attempts were allowed, and whether the model was fine-tuned for specific tasks. 3. Check whether the repository contains weights, not just code. This is essential for reproducing the result. 4. Look at the compute requirements for inference. World models are heavier than direct policies, and control frequency is critical for a manipulator. 5. Find out how the post-deployment self-learning mechanism works, if there is one: what reward signals it uses and who labels success.

Open code makes LoopWAM a convenient baseline for labs building their own world models. If the weights and evaluation protocol turn out to be complete, the project could become a reference for subsequent work on RoboChallenge. If not, it will be one more strong result that is hard to verify. It is too early to judge: the numbers from the paper are needed.