https://arxiv.org/html/2604.27792v1

License:
arXiv.org perpetual non-exclusive license

arXiv:2604.27792v1 [cs.RO] 30 Apr 2026

MotuBrain: An Advanced World Action Model for Robot
Control

MotuBrain Team

https://www.shengshu.com/en/motubrain

1Introduction

Recent progress in embodied foundation models has been
driven by Vision-Language-Action (VLA) policies [20, 6, 26, 22, 38, 10, 33, 8], which map visual
observations and language instructions in pretrained Vision-Language Models
(VLMs) [4, 1] directly to
robot actions. Inheriting rich semantic priors, VLAs generalize well across
diverse objects and language instructions, achieving strong performance on a
wide range of robotic tasks. However, since they are primarily pretrained on
static image–text data, they often neglect the perception and prediction of
fine-grained world dynamics that are essential for precise robotic control,
leading to a superficial imitation of behaviors rather than temporal
understanding of the physics of the world.

With the rise of video generation models [34, 3, 39, 30, 23], a growing body of
research has begun exploring how to adapt these models for world
modeling [15, 17, 28, 9, 18, 16]. A world model
aims to predict how the environment evolves in response to actions—a capability
that aligns directly with what a video generation model does when conditioned
on past observations and actions, namely forecasting future visual states. Conceptually,
video generation models are naturally suited to this task, as video generation
models are pretrained on vast and diverse web video data, which equips them
with rich spatiotemporal priors regarding object permanence, physical dynamics,
and human-object interactions. This enables them to generalize more effectively
and reason about novel scenarios where in-domain data are scarce. As a result,
leveraging video generation models as a foundation for world modeling has
emerged as a promising paradigm for scalable robot policy learning.

Following this insight, early attempts largely followed a
two-stage paradigm: Video Generation Model (VGM) plus Inverse Dynamics Model
(IDM) [14, 19, 40, 29, 13]. In this
framework, a video diffusion model pretrained on large-scale web data first
predicts future visual trajectories from current observations and language
instructions. Subsequently, an inverse dynamics model infers actions from the
generated future frames. While this paradigm successfully leverages rich
spatiotemporal priors from video data to achieve broad generalization, it
suffers from a critical drawback in that errors in video prediction accumulate
over time, which leads to compromised action accuracy and downstream policy
performance.

To mitigate this issue, a subsequent line of research has
explored World Action Models (WAMs) that unify visual dynamics and action
prediction within a single, jointly optimized objective [5, 37, 36, 21, 25, 41, 24, 27]. Unlike the
VGM+IDM pipeline, which decouples forecasting from action inference, WAMs
simultaneously predict future visual states and actions in an aligned manner.
This integration offers two key advantages: (1) it avoids the cascading errors
inherent in sequential video prediction, and (2) it overcomes the fragmented
functionality of conventional embodied systems, where semantic understanding,
dynamics modeling, and action generation are typically learned from disparate
supervision sources.

In our view, the core source of intelligence in such a
unified model is its ability to absorb large-scale heterogeneous multimodal
data under one unified training recipe. In principle, this includes pure video
data without action annotations, robot data with aligned video-language-action
trajectories across different embodiments, and even task-agnostic interaction
data with partially missing modalities [32]. By contrast, VLA
learning primarily relies on robot task trajectories with aligned
observation-language-action supervision, and adaptation to the target robot is
primarily coupled to embodiment-specific action data [22, 6, 12].

Motus [5] was an early
step in this direction. It established a unified world-action formulation in
which video and action are modeled in a shared generative framework, so that
policy modeling, world modeling, video generation, inverse dynamics, and joint
video-action prediction become different inference modes of the same model. It
also showed that, with UniDiffuser-style continuous multimodal modeling and a
Mixture-of-Transformers design, a world action model can absorb heterogeneous
multimodal data rather than being restricted to embodiment-specific task
trajectories.

Building on this foundation, we present MotuBrain. Like
Motus, MotuBrain adopts UniDiffuser [2] to jointly
model and schedule the two continuous modalities, namely video and action, and
uses a three-stream Mixture-of-Transformers architecture to integrate video
generation, action modeling, and language conditioning within one system. This
unified formulation again supports inference over five distributions with the
same model: vision-language-action policy modeling, world modeling, video
generation, inverse dynamics, and joint video-action prediction. More
importantly, it preserves the key advantage of unified world-action modeling:
the model can learn from a much broader family of multimodal data, including
video-only data without action labels, interaction data without explicit task
language, and heterogeneous robot trajectories collected from different
embodiments.

MotuBrain further extends this paradigm in several
practically important directions. It introduces a unified multiview
representation that supports an arbitrary number of camera views under
different camera layouts, rather than depending on a fixed visual input format.
It uses an independent text stream to more tightly couple high-level semantics
with low-level control, making instruction following an explicit part of action
generation and improving semantic understanding. It adopts a unified action
representation across embodiments, enabling the model to capture transferable
control regularities rather than overfitting to the action format of a single
robot. Beyond architecture and pre-training, MotuBrain also involves
a post-training and deployment recipe tailored to long-horizon real-world
control: autoregressive diffusion rollout enables temporally extended
execution, V2A-style asymmetric dependency enables action-only inference
without explicitly generating future video, and real-time chunked closed-loop
execution reduces boundary discontinuities during asynchronous control.
Finally, we develop a systems-oriented inference stack including denoising-step
reduction, compilation with CUDA-graph-friendly execution, FP8 quantization,
and DiT caching, which together deliver more than  end-to-end speedup
over the naive baseline and make large world action models practical for
real-time robotic deployment.

Figure
1:Overview of MotuBrain’s architecture. MotuBrain builds on a
unified video-action backbone and adopts a three-stream Mixture-of-Transformers
architecture with text, video, and action streams. It further uses an H-bridge
attention design to balance cross-modal interaction and efficiency, while supporting
flexible multiview inputs through view-dependent 3D RoPE offsets.

2Method

In this section, we present the overall methodology
of MotuBrain from four aspects: model architecture, pre-training,
post-training, and inference. We first introduce the core architectural design
and the main acceleration techniques integrated into MotuBrain in
Section 2.1.
We then describe the pretrained foundation model used in our system in
Section 2.2.
After that, we present the post-training pipeline, including sparse adaptation
and step distillation, in Section 2.3.
Finally, we describe the inference-time acceleration strategies deployed
in MotuBrain in Section 2.4.

2.1Model Architecture

MotuBrain first adopts UniDiffuser [2] to jointly
model and schedule the two continuous modalities, i.e., video and action, so
that all interaction patterns between them are captured in a unified generative
framework. As a result, a single training procedure supports inference over five
distributions: vision-language-action policy modeling, world modeling, video
generation, inverse dynamics, and joint video-action prediction, which are
formulated in Table 1.

Table 1:Different prediction modes (non-autoregressive
mode).

On top of this unified video-action backbone, MotuBrain introduces
a three-stream Mixture-of-Transformers (MoT) architecture with a text stream, a
video stream, and an action stream. The text stream serves as a conditioning
branch: its hidden states participate in Transformer attention, but no output
head is applied to text tokens. Maintaining a dedicated text stream also
improves semantic understanding and instruction-following capability. The video
and action streams are both trained with flow matching, predicting the velocity
fields of video latents and action tokens, respectively.

The inputs consist of text tokens, condition-image latents
encoded by a Vidu VAE, noisy future video latents, and noisy action tokens. The
condition image is represented as the first video latent frame and is
teacher-forced in the video stream, while the remaining video latents and
action tokens are denoised by the video and action streams. Cross-modal
interaction is implemented through joint attention over video, action, and text
tokens.

Instead of using full video-action joint attention in all
layers, we adopt an H-bridge design following HBridge [35]. Concretely, full
V-A joint attention is applied only in the middle 50% of Transformer layers,
while the bottom and top 25% layers use decoupled attention. In the decoupled
layers, video tokens and action tokens are processed independently and do not perform
joint-attention. This design reduces the cost of dense cross-modal attention,
improves efficiency, preserves modality-specific representations in shallow and
deep layers, and avoids injecting excessive modality-irrelevant information
into every layer, while still allowing semantic alignment and policy grounding
in intermediate layers.

For multiview inputs, each view is encoded independently by
the Vidu VAE and concatenated at the token level. Leveraging 3D RoPE, we
introduce view-dependent offsets only along the spatial dimensions while
keeping the temporal dimension unchanged. This effectively maps different views
to distinct regions in a shared spatial positional space, enabling seamless
support for an arbitrary number of camera views without modifying the backbone
architecture.

2.2Pre-training

Our pre-training data follows a four-level data
pyramid [5],
organized from broad visual diversity to embodiment-specific control signals:
Internet videos, ego-centric videos, heterogeneous-embodiment data, and
specific-embodiment data. The main motivation is to maximize data efficiency.
The bottom level uses large-scale Internet videos to train the video generation
model Vidu [3],
which serves as the foundation model of MotuBrain. The second level
introduces ego-centric videos, which provide first-person interaction patterns
and hand-object dynamics that are closer to embodied manipulation. The third
level uses heterogeneous-embodiment data collected from different robot platforms,
tasks, and scenes. In our setting, we only use dual-arm robot data at this
level. The top level consists of specific-embodiment data collected on the
target robot configuration, which further aligns the model with the final
action space, kinematics, camera setup, and deployment distribution. This
hierarchy reflects a central design principle of MotuBrain: the model
should learn from heterogeneous multimodal data wherever available, rather than
restricting supervision to a single narrow data format.

Starting from the pretrained Vidu weights [3], we perform
two-stage pre-training corresponding to the second and third levels of the data
pyramid. In the first stage, we train only the video branch on ego-centric and
heterogeneous-embodiment data, while keeping the randomly initialized action
branch unchanged. Accordingly, the optimization objective in this stage
contains only the video loss. This stage is designed to adapt the
Internet-scale video prior to embodied manipulation and to obtain a video world
model that can understand and predict bimanual interaction dynamics. To improve
robustness to imperfect visual conditioning, we follow the noisy-conditioning
strategy of LingBot-VA [24] throughout
training, including stage 1, stage 2, and non-autoregressive post-training, but
not the autoregressive policy setting. In addition, for multiview data, we
randomly drop auxiliary views with probability  during pre-training,
so that the model can better adapt to varying numbers of camera views and
imperfect visual observations. Concretely, with probability , we perturb the
conditioned-frame latent as

(1)

and leave it unchanged otherwise. In the second stage, we
initialize from the first-stage checkpoint and train only the action branch on
heterogeneous-embodiment data, while freezing the video branch. At this stage,
we use a unified action representation across embodiments.

Concretely, let the absolute end-effector chunk be  and let  denote the
end-effector state of the conditioned frame. We define the corresponding
relative chunk as , where each action is
represented as

(2)

Here  denotes a
component-wise pose difference: the position is computed by direct subtraction,
the rotation is computed by composition with the inverse reference rotation,
and the gripper state is kept unchanged. Writing  with position , rotation , and gripper state , we have

(3)

and

(4)

The raw input pose is provided in quaternion format, while
the training target uses a D rotation representation.
Each end-effector action therefore has dimension , consisting of position,
rotation, and gripper state. We only normalize the gripper dimension to  and keep the remaining
dimensions in their original physical scales. Using relative end-effector
coordinates with respect to the conditioned frame improves compatibility across
different robot embodiments and initial poses, making the action space more
consistent under heterogeneous pre-training and allowing the target embodiment
to be adapted with less embodiment-specific data. Although only the action
branch is updated, stage 2 still optimizes both video and action objectives
under the unified formulation. We use separate SNR-based timestep sampling for
the two modalities, with  for video and  for action. The
overall stage-2 objective is the weighted sum of the video and action losses,

(5)

where

(6)

This design encourages alignment between the action and
video modalities within the unified model: action learning benefits from the
visual dynamics encoded by the video branch, while the action-conditioned
interaction also improves video prediction by injecting control-relevant
information into the shared modeling process.

Figure 2:Attention masks used in all training stages and
post-training modes. (a) Stage-1 pre-training updates only the video branch.
(b) Stage-2 pre-training uses full joint attention over language, conditioned
video, video, and action tokens. (c) Non-autoregressive post-training disables
video-to-action attention. (d) Autoregressive post-training applies a causal
mask over temporally ordered video and action tokens.

2.3Post-training

While pre-training aims to build a unified world action
model from broad and heterogeneous data, post-training focuses on adapting the
model to the target embodiment. We start from the checkpoint obtained after the
second stage of pre-training and finetune it on specific-embodiment data. In
this stage, we consider two post-training settings, namely a non-autoregressive
(Non-AR) setting and an autoregressive (AR) setting, and train separately under
each setting. The attention masks used in stage 1, stage 2, Non-AR, and AR are
summarized in Figure 2.
The two post-training settings mainly differ in sequence factorization and
attention masking.

Under the Non-AR setting, the model denoises all video and
action tokens within the full observation window in a single forward pass,
where  denotes the
observation at step ,  is its VAE latent, the
VAE temporally compresses every  consecutive video
frames into a single latent frame, and the action stream operates at  actions per raw video
frame. Each latent frame therefore corresponds to  action tokens.
Consequently, predicting  future video latents
amounts to generating  actions, enabling
high-frequency control while keeping the visual stream compact. As noted in
Section 2.2,
we continue to apply the same LingBot-VA-style noisy-conditioning augmentation
to the conditioned-frame latent in Non-AR post-training. This improves
robustness to noisy conditioning and helps the model recover from partially
corrupted observations.

For long horizon tasks, we adopt an AR formulation based on
chunk-level factorization, as illustrated in Figure 2(d).
We partition each episode into non-overlapping chunks and process all chunks in
parallel during training with a block-causal attention mask. The input sequence
contains language tokens, conditioned-image tokens, clean video tokens from
teacher-forced past observations, and the noisy video and noisy action tokens
of the current target chunk. Unlike Non-AR post-training, the AR setting does
not apply the noisy-conditioning augmentation to the conditioned frame. For
chunk , tokens can attend only to
clean visual context from preceding chunks, but never to future chunks.
Importantly, our AR model does not involve clean action tokens, because doing
so would break the unified relative-EEF action representation. At deployment,
the model rolls out sequentially, using the newly observed frame as clean
context for the next chunk.

To better align post-training with deployment, we adopt
V2A-style attention in both operating modes. Under this attention pattern,
action tokens attend to video and language tokens, while video tokens never
attend to action tokens. Combined with the UniDiffuser formulation, where video
and action are sampled with independent timesteps, this asymmetric dependency
makes it possible to use an action-only suffix during inference: after a short
joint denoising prefix, the video stream can be frozen and only the action
stream continues to update while attending to the cached visual-language
context. This design is motivated by efficiency rather than a change in
training objective, and in practice it enables substantial inference
acceleration without degrading task success rate.

2.4Inference

2.4.1Inference Optimization

A practical concern with world-action models is inference
latency: jointly denoising a high-dimensional video latent and an action latent
over many diffusion steps couples two of the slowest computations in modern
generative modeling, and naively deployed WAMs typically run far below the
control rates required for high frequency manipulation. We adopt and extend a
stack of optimizations to tackle this problem. Crucially, we verify on the
RoboTwin2.0 that this speedup is essentially lossless: average task success
rates fluctuate within sub-percent margins across the optimized and unoptimized
configurations, indicating that the gains come from removing genuinely
redundant computation rather than from sacrificing model fidelity. We detail
each component below.

Noise sampling.

We use separate SNR-based timestep sampling for the two
modalities during training. Specifically, we set  for video and  for action. This makes
video timesteps more likely to be sampled from noisier regions, while action
timesteps are sampled more uniformly. As a result, the model becomes more
robust at predicting accurate actions under noisy visual conditions. This improves
both robustness and convergence, allowing the number of inference steps to be
reduced from 50 to 30 without performance degradation.

Compiling.

We optimize the inference graph and fuse the operators
with torch.compile to reduce the overhead of repeated denoising.
Since the model is rewritten as a single-GPU, pure-PyTorch inference model, the
main Transformer computation can be compiled directly at inference time. This
optimization primarily improves execution efficiency for the repeated DiT
forwards used during sampling.

DiT Cache.

We adopt a DreamZero-style DiT caching strategy [36] to exploit
temporal redundancy across denoising steps. Let  denote the velocity
prediction at denoising step . We measure the similarity
between two consecutive predictions by

(7)

When , where  is a predefined
threshold, we skip a small number of subsequent DiT evaluations and approximate
the skipped predictions from recent history, i.e.,

(8)

where  is the cache length.
The cache can be applied to either the video velocity or, in action-only mode,
the action velocity, and is reset for each inference call or chunk.

FP8 Quantization.

We further reduce DiT inference cost with FP8 quantization.
The implementation replaces eligible nn.Linear layers with FP8 linear
layers. We store the weights in float8_e4m3fn with a per-tensor
scale, dynamically quantize activations to FP8 at runtime, and compute matrix
multiplications through torch._scaled_mm, returning outputs in the
original compute dtype. Layers whose input or output dimensions are not
divisible by 16 are skipped to satisfy the kernel alignment requirement. This
optimization mainly targets the large linear projections in attention and MLP
blocks, reducing memory bandwidth and GEMM cost on FP8-capable GPUs. Since
quantization is applied after checkpoint loading and before compilation, the
compiled graph traces the quantized linear operators directly.

V2A-Style Inference.

For models trained with V2A attention, the video stream is
prevented from attending to action tokens, while the action stream can still
attend to visual and language context. This asymmetric dependency enables an
efficient inference schedule. Instead of denoising video and action jointly for
all sampling steps, we use a short joint denoising prefix followed by an
action-only suffix:

(9)

where  and  denote the video and
action latents at denoising step , respectively. After
step , the video latent is
frozen, and the video-text branch is executed once to build per-layer cached
keys and values for the fixed visual-language context. Subsequent denoising
steps update only the action tokens: action queries attend to the cached video/text
keys and values together with their own action keys and values. This removes
the repeated video-stream computation from the latter part of sampling while
preserving the same attention semantics as the V2A-style model. Combined with
the other inference optimizations above, this design enables

ours to reach a 11 Hz inference frequency, exceeding typical human reaction
speed.

Action Smoothing.

Following DreamZero [36], MotuBrain applies
action chunk smoothing to improve execution smoothness. Each chunk is upsampled
to twice its original temporal resolution, smoothed with a Savitzky-Golay
filter, and then downsampled back to the original resolution.

Frequency-aware interpolation.

After smoothing, the action sequence is interpolated
according to the ratio between the model action frequency and the low-level
control frequency. This frequency-aware interpolation preserves the temporal
scale of the model prediction: rapid transitions remain fast during execution,
while slower motion segments are expanded over a longer control duration. For
manipulation tasks with different motion phases, this helps preserve the
velocity profile predicted by the model.

Table 2:Cumulative speedup from inference optimizations.
Each row applies the listed technique on top of all preceding ones. Latency is
measured end-to-end on the non-autoregressive model. Per-step latency is
reported where applicable.

2.4.2Real-time Inference and Execution

Real-time control is essential for robotic deployment, while
the inference latency of world action models is usually non-negligible.
MotuBrain therefore decouples the model inference loop from the robot action
execution loop: the controller executes the current action chunk at the target
control frequency, while the world action model asynchronously generates the
next chunk from the latest observation.

However, directly switching to the newly generated chunk may
introduce chunk-boundary discontinuities, such as action regression, velocity
jumps, and high-frequency jitter, since adjacent chunks may be generated from
different observations and action modes. To reduce such boundary mismatch, we
adopt an RTC-inspired strategy [7]. The unexecuted part
of the current chunk is used as a constraint for the next generation and fused
before denoising, where the latency-affected prefix is treated as a frozen
region and the remaining overlapping actions are used as soft constraints.

Let the current action chunk be

(10)

where  denotes the prediction
horizon. If a new inference request is launched after  actions have been
executed, the remaining actions in the current chunk are

(11)

Given the end-to-end inference latency  and the controller
period , the inference delay
measured in action steps is defined as

(12)

The first  steps of the next
chunk correspond to the period in which the previous chunk is still being
executed while the new inference is running. These actions are therefore fully
constrained by the remaining actions from the previous chunk. For the following
overlapping region, we apply a smooth decay weight to gradually reduce the
influence of the previous chunk and allow the new prediction to take over.

Let  denote the normalized
progress within the fusion window, where  is the action index in
the new chunk, and  denotes the end of the
fusion window. The exponential decay function used in our implementation is

(13)

Since  increases from 0 to 1,
the fusion weight is defined as

(14)

The fused action is computed as

(15)

To improve robustness under variable inference and
communication latency, MotuBrain maintains a delay queue  that stores recent
inference delays, and the system uses  as a conservative
estimate for the next inference request. This estimated delay determines the
length of the frozen prefix. The fusion window is also adjusted accordingly:
when the estimated delay increases, more steps are treated as fully constrained;
when the estimated delay decreases, more future actions are left to be updated
by the new prediction. This makes the asynchronous execution more stable under
fluctuating network and model latency.

3Evaluations

In this section, we evaluate MotuBrain in both
simulation and real world environment.

3.1Evaluation on Simulation Environment

Following the protocol of RoboTwin 2.0 [11], we adopt a
multi-task training setup where all models are trained with 2,500
demonstrations collected from clean scenes (50 per task) together with 25,000
demonstrations from heavily randomized scenes (500 per task). For training
efficiency and temporal consistency across tasks, we downsample videos to 5 Hz
and action sequences to 10 Hz. We further conduct ablation studies over
multiple policy architectures to analyze how different design choices affect
the performance of MotuBrain under the same data and evaluation
setting.

Table 3:Robotwin 2.0 Results. Following previous works,
MotuBrain is fine-tuned from pre-trained weights on the official RoboTwin 2.0
dataset (clean + randomized), yielding the evaluation results presented in the
table. MotuBrain-Non-AR represents non-autoregressive mode.

Table 4:Per-task success rates on RoboTwin under clean and
randomized evaluation settings.

As shown in Table 3, MotuBrain achieves
the best average success rate on RoboTwin, reaching 95.8% in clean scenes and
96.1% in randomized scenes. It ranks first in both settings and is the only
model on the leaderboard whose average score exceeds 95% under randomized
evaluation. At the per-task level, as summarized in Table 4,
MotuBrain attains a perfect score on 24 tasks under the clean setting and 25
tasks under the randomized setting, with 19 tasks achieving 100% success in
both settings. Moreover, it surpasses 90% success on 42 clean tasks and 44
randomized tasks, demonstrating consistently strong generalization across the
full 50-task benchmark and strong ability to act robustly in the world.

From the detailed breakdown in Table 4,
the largest gains are concentrated on tasks that require robust multi-stage
manipulation and precise interaction under visual variation. In particular, we
observe clear improvements on handover and coordination-heavy tasks such
as Handover Block, articulated-object interaction tasks such
as Open Microwave, Press Stapler, and Turn
Switch, as well as fine-grained spatial arrangement tasks including Blocks
Ranking Size, Move Can Pot, Place A2B Left,
and Place Can Basket. These results suggest
that MotuBrain is especially effective on tasks demanding accurate
temporal modeling, stable contact-rich control, and robustness to scene
randomization.

We further observe a favorable multi-task scaling trend: as
the number of training tasks increases, the average success rate
of MotuBrain continues to improve, suggesting that more tasks expose
the model to more reusable world knowledge about objects, contacts, and
temporal transitions, as illustrated in Figure 3.
In contrast, conventional VLA baselines exhibit noticeably weaker scaling in
the same regime, indicating stronger task interference and less effective
knowledge sharing across tasks. We also find that MotuBrain is
substantially more data-efficient than both conventional VLA baselines and the
previous Motus model: under the same data scaling protocol, MotuBrain reaches
stronger performance with fewer training trajectories and continues to improve
more effectively as data grows, as shown in Figure 4.

Taken together, Figures 3 and 4 further
suggest that, in our setting, increasing task diversity is more effective than
merely scaling up the amount of data collected for a fixed task set, as
evidenced by the steeper improvement trend in the task-scaling curve. This
observation is consistent with the hypothesis that broader task coverage
exposes the model to a richer set of interaction patterns, object affordances,
and temporal transitions, thereby improving knowledge reuse and cross-task
generalization more efficiently than data duplication alone.

Figure
3:Task scaling results. For each point on the curve, we train the model using
data from only the specified number of tasks and keep optimization running
until convergence, i.e., until the average success rate becomes approximately
stable. As the number of training tasks increases, MotuBrain shows a
clear upward trend in average success rate and consistently stronger scaling
behavior than conventional VLA baselines.

Figure
4:Data scaling results. For each data budget, we uniformly subsample
demonstrations from every task to keep the per-task data distribution balanced.
The number of training steps is set proportional to the total amount of
training data; in particular, training on the full 27,500-trajectory dataset
uses 50,000 optimization steps. MotuBrain continues to benefit from
additional training data, indicating that larger-scale supervision improves
policy performance and robustness.

3.2Evaluation on World Models

Beyond action execution, we also evaluate
whether MotuBrain can understand and predict changes in the world
itself. We adopt WorldArena [31], a unified
benchmark that scores embodied world models through 16 numerical metrics across
six perceptual sub-dimensions—visual quality, motion quality, content
consistency, physics adherence, 3D accuracy, and controllability. We use the
entire 2,500-trajectory release for high-resolution training and report
benchmark submission results. Because MotuBrain is a unified
world-action model, the same backbone simultaneously serves as the manipulation
policy used in Section 3.1 and
as the world predictor evaluated here. For WorldArena, we run the model in its
forward-dynamics mode (FDM), which serves as an action-conditioned world model.
We train and run inference at high resolution with a  Hz video stream paired with  Hz actions ( ratio), using
classifier-free guidance to improve instruction-following capability.

As shown in Figure 5, MotuBrain attains
the highest EWMScore (63.77) among embodied world models and video
baselines on the WorldArena leaderboard. Its margin over the strongest
video-generation baseline shown here (Wan2.6, ) is approximately four
points. Notably, MotuBrain ranks first on motion quality and remains
competitive on several other dimensions, indicating that strong dynamics
modeling can be achieved without sacrificing overall perceptual quality.

The detailed per-metric breakdown in Table 5 shows
that MotuBrain’s lead is driven primarily by its strong performance in
the Motion Quality dimension. The Flow Score, which measures
the overall magnitude of motion in the predicted video, is the highest among
all baselines, indicating that MotuBrain generates sequences with
sustained inter-frame displacement rather than near-static clips that collapse
to a stable but trivial solution. The leading Motion Smoothness indicates that
the produced motion unfolds with realistic inertia rather than frame-to-frame
jitter. The high Dynamic Degree, which measures whether motion concentrates in
the regions that actually matter, further confirms that MotuBrain’s
predicted activity is localized on the embodiment and the manipulated objects
rather than diffused into background flicker. Together, these three metrics
show that MotuBrain predicts motion that is substantial, smooth, and
concentrated in embodiment-relevant regions.

This distinction matters because the original WorldArena
study reports that EWMScore correlates only weakly with downstream
action-planning success (), reflecting the well-known
perception–functionality gap: visually impressive world models often fail when
used for control, whereas functionally useful ones often appear visually
unpolished. MotuBrain helps bridge this gap. On the perceptual side,
it ranks first on the EWMScore leaderboard among embodied world models and
video baselines; on the functional side, it achieves an average success rate
of  across the 50 RoboTwin
manipulation tasks (Section 3.1),
the highest among all evaluated VLA and world-model baselines. This
cross-benchmark consistency suggests that the dynamics
representation MotuBrain learns is simultaneously perceptually
faithful and functionally actionable when consumed by a policy.

Table 5:Per-metric WorldArena results and overall EWMScore
comparison. We compare MotuBrain with several publicly listed and
representative entries on the WorldArena leaderboard. The best score in each
row is shown in bold. EWMScore is the arithmetic mean of all 16 normalized
metrics, scaled to .

Figure
5:WorldArena public leaderboard.

3.3Real-World Experiments

We further investigate rapid deployment to new embodiments
in real-world settings. Starting from a pretrained model, MotuBrain can
be adapted to a new embodiment using only 50–100 same-embodiment trajectories.
We validate this capability across multiple humanoid platforms. Notably, these
results are achieved without relying on auxiliary components such as VLM-based
planners, dual-system decompositions, external memory modules, or additional
reinforcement/retry data. In other words, the native world action model alone
provides a sufficiently strong control prior to enable practical transfer.

3.3.1Real-World Performance

Quantitative Evaluation.

We conduct a quantitative evaluation
of MotuBrain on a suite of representative real-world household tasks.
Performance is measured using a normalized scoring metric with a maximum value
of 100, where each sub-task step contributes equally to the overall score. A
step receives full credit if it is successfully completed on the first attempt;
if it succeeds after one retry, two retries, or three or more retries, it
receives 80%, 50%, or 0 credit, respectively.

For each task, the model is trained on 100 task-specific
trajectories consisting of single, coherent, and successfully executed compound
actions, without atomic-level annotations. We further assess performance across
tasks with varying execution horizons and differing numbers of atomic actions.
The quantitative results are summarized in Table 6,
Table 7,
and Table 8.

In the Making Oden task, which requires simultaneous
bimanual manipulation with one arm pouring a drink and the other scooping
dumplings, MotuBrain is evaluated over five consecutive trials, with
an average execution time of 33 seconds across seven atomic actions. The robot
executes the subtasks concurrently with its left and right hands, achieving an
overall score of 98.54, with only a single atomic action failure observed.

Table 6:Per-step performance evaluation of the bimanual
Making Oden task. The overall score reaches 98.54, with only a slight
deficiency in the pouring step.

In the Mixing Cocktails task, which involves an extended
long-horizon sequence with complex multi-step dependencies across 15 atomic
actions, the model is evaluated over an average execution time of 124 seconds.
Across seven consecutive trials, MotuBrain attains an overall score
of 97.34, demonstrating stable performance under substantially increased
temporal complexity.

Table 7:Per-step real-world evaluation results for the
long-horizon Mixing Cocktails task. Each sub-task contributes equally to the
final score, and MotuBrain achieves strong overall performance with
only minor deficiencies in a small number of steps.