How Muka Robotics’ LJM Reached No. 2 on WorldArena With 32 GPUs

Muka Robotics’ LJM, short for Latent Joint-conditional Model, stands out not because it proves that small-scale compute will always beat much larger training runs. Its significance is the different way it allocates learning capacity: it separates reasoning about what an action will do in the physical world from rendering that outcome as a coherent video, while keeping the two processes tightly connected.

According to public reports, Muka submitted LJM to WorldArena under the anonymous name SisyphusWorld. It reportedly achieved an EWMScore_P of 73.06, ranking second on the public leaderboard behind Xiaomi UNIS at 73.64. The same reports list sub-scores of 89.17 for Motion Quality, 92.60 for 3D Accuracy and 85.93 for Controllability, and say the model was trained with 32 Zhenwu 810E GPUs. 1

Those results are notable, but the currently available evidence is largely based on platform and media reports. It is not a substitute for a fully public, independently reproduced evaluation or a detailed ablation study.

Start with the interaction, then generate the image

Conventional action-conditioned video diffusion models often feed robot actions into a video-generation network as relatively low-dimensional conditions. That can make the generated sequence respond to the action, but the model still has to infer what the action values mean through video- or pixel-level training objectives: whether an arm has contacted an object, whether a gripper has grasped it, where the object will move and whether spatial relationships remain consistent.

LJM aims to make that intermediate step explicit. It converts a language instruction, visual history, an anchor observation and planned future actions into learnable interaction latents. Those latents then constrain the generation of future video. 1

The point is not that video generation becomes less important. Rather, LJM’s premise is that a video generator can devote more of its capacity to visual continuity and detail when a dedicated representation layer is responsible for predicting robot-object interactions.

A two-expert division of labour

LJM is described as two cooperating experts:

• Latent Reasoning Expert: Built on Qwen3-VL-2B, this component reads the instruction, visual sequence, anchor observation and planned actions, producing continuous reasoning tokens. “Reasoning” here does not mean a natural-language chain of thought. It refers to latent-space representations of the robot’s expected state, likely scene changes and action-driven motion. 1
• World Modeling Expert: Built on the Wan2.2-TI2V-5B Diffusion Transformer, this component generates future video in VAE latent space, translating the first expert’s interaction representation into a visually continuous sequence. 1

A useful analogy is a production pipeline with an interaction script and a visual renderer. The first describes the consequences of an action; the second depicts those consequences as a plausible visual future. That division offers a reasonable architectural explanation for strong reported motion, spatial and controllability results under limited training resources—but it is still an explanation, not conclusive proof of causation.

Three physical prediction targets help ground the latent space

Adding a latent variable alone does not ensure it represents physical processes. LJM also reportedly applies three prediction constraints tied to observable data:

1. 3D end-effector position, linking the representation to the robot arm’s spatial trajectory;
2. Visual dynamics of scene objects, encouraging the model to track whether manipulated objects change in ways consistent with the action;
3. Optical-flow direction, making the model account for the direction of movement in the image. 1

These objectives mean the latent representation is not trained solely to help generate convincing-looking frames. It must also account for where the robot is going, which objects are changing and which way visible motion is moving. Public descriptions characterize this as interleaved supervision over future robot state, optical flow and visual latents. 1

That does not guarantee rigorous physical understanding in every contact, friction or occlusion scenario. But it gives training a more direct interaction signal than video reconstruction alone.

Why shared attention may matter

LJM also uses shared attention in a Mixture-of-Transformers design. Rather than injecting one fixed conditioning vector into a video generator once, reasoning tokens and video tokens exchange information across model layers. 1

The intended result is ongoing two-way coordination:

• Interaction predictions supply spatial and motion constraints for video generation.
• Context from the evolving video representation can help update the interaction representation.

Compared with one-way, fixed conditioning, layer-by-layer exchange is better suited in principle to a long process where object position, occlusion or motion changes as the future sequence unfolds. Public materials describe this mechanism, but do not provide independent ablations sufficient to quantify its individual contribution. 1

What “32 GPUs” does—and does not—mean

Reports say LJM used 32 Zhenwu 810E GPUs across DROID pretraining and RoboTwin continued training. 1 The more meaningful takeaway is not that video world models have suddenly become cheap. It is that the architecture may reduce the burden on the generator to discover interaction structure indirectly from pixels.

Put differently, the efficiency claim rests on inductive bias: allocating learning capacity to a latent representation with explicit interaction meaning and observable constraints, rather than asking one monolithic video model to learn action control, spatial structure, object dynamics and image quality all at once.

GPU count alone, however, cannot establish a like-for-like comparison of total training compute. Training duration, dataset size, parameter count, resolution, parallel efficiency and training stages can all vary substantially between systems. Available information is therefore insufficient to calculate how much total compute LJM saved relative to competing models.

What is known about Muka Robotics

Public reporting says Muka Robotics was established in April 2026 and focuses on world models for real-world robotics. One report identifies Hong Kong University of Science and Technology PhD Chi Xiaowei as the founder, legal representative and actual controller; another describes him as co-founder and CEO. 1838

The available materials do not provide sufficiently reliable and consistent detail to confirm the full team’s previous companies, laboratories or institutions. The cautious conclusion is that Chi’s HKUST doctoral background is publicly reported, while broader claims about the team’s origins should not be inferred from that alone. 1838

What the result actually suggests

SisyphusWorld’s leaderboard performance supports a hypothesis worth testing further: robot world models may not need to rely only on scaling video generators to pursue realism. They may also benefit from treating interaction consequences as an explicit learning target and coupling that representation closely to video generation.

A leaderboard rank alone does not show that explicit physical reasoning is universally better than scaling training, nor does it establish that LJM will outperform alternatives across robots, data distributions or real-world deployments. Testing those claims would require fuller public technical documentation, including ablations of the physical-supervision targets, shared attention and dual-expert split, plus evaluations across datasets, robot embodiments and closed-loop real-world control.

For now, the most defensible conclusion is that LJM presents a compelling engineering proposition: train interaction reasoning and visual generation together, without treating them as the same task, and it may be possible to improve both physical consistency and visual quality with more focused compute. 1