Awesome-WAMs
GitHub
PAPER REPORTENAll readings ↗

Motus: A Unified Latent Action World Model

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Hongzhe Bi; Hengkai Tan; Shenghao Xie; Zeyuan Wang; Shuhe Huang; Haitian Liu; Ruowen Zhao; Yao Feng; Chendong Xiang; Yinze Rong; Hongyan Zhao; Hanyu Liu; Zhizhong Su; Lei Ma; Hang Su; Jun Zhu

Affiliations: Dept. of Comp. Sci. and Tech., Institute for AI, BNRist Center, THBI Lab, Tsinghua-Bosch Joint ML Center, Tsinghua University; Peking University; Horizon Robotics

Source: CVPR 2026 · ref-fa771119edf30fd77084 ↗ · Catalog record

Reading: 274 / 558 · 6 original figures & tables · ~20 min ·

1. Paper overview

In one sentence: Motus unifies video and action generation through interacting experts and motion-based pretraining, with stronger average robot performance but unresolved attribution and reporting gaps. e01e03e04e05e11e12e13e14

At a glanceWhat to know
Research problem
Source description

Robot policies need language understanding, visual dynamics and executable controls, yet heterogeneous videos lack compatible action labels. Motus addresses both unified prediction and cross-embodiment motion transfer. Its manipulation formulation predicts action chunks from the current image and instruction, optionally proprioception. e02e03

Core mechanism
Source description

A Mixture-of-Transformers (MoT) network preserves expert-specific transformations while exchanging features through Tri-model Joint Attention. Separate video/action noise levels enable conditional, marginal and joint generation. e03e07e08

A key reported resultRoboTwin 2.0 multi-task manipulation: Motus: 88.66 clean; 87.02 randomized.

Average task success (%). 50 tasks; 2,500 clean plus 25,000 randomized demonstrations; 40k fine-tuning steps; 100 execution trials per task.

X-VLA: 72.80/72.84; π0.5: 42.98/43.84 (clean/randomized). Calculated randomized gains are 14.18 and 43.18 percentage points. The prose claim of over 45 points over π0.5 in randomized scenes conflicts with these means; the clean gain is 45.68 points. Uncertainty is unreported. e09e11

Reading caution
Reader analysis

Figure 6/Table 14 show clean no-pretraining/Stage-1 means 77.56/82.26; Table 2 prints 72.8/82.86. Table 3 gives AC-One cube baseline 46, whereas Table 15 gives 57.5. These conflicts remain unresolved. e09e11e12e13e14

Core contributions

  • Source description

    A Mixture-of-Transformers (MoT) network preserves expert-specific transformations while exchanging features through Tri-model Joint Attention. Separate video/action noise levels enable conditional, marginal and joint generation. e03e07e08

  • Source description

    A flow-reconstruction bottleneck with weak action supervision enables a three-stage training recipe spanning six data levels, from inherited web priors to target-robot demonstrations. e04e05

Figure 1. Separate expert transformations exchange information through Tri-model Joint Attention. Original paper, p. 2 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read upward from the inputs. At right, an image and instruction enter the pretrained VLM; its snowflake matches the frozen VLM described in Stage 2. The resulting features enter the understanding expert. At left and center, video and action encoders supply tokens to their own experts. The shared horizontal attention block lets the branches interact before their separate feed-forward transformations and output decoders. The arrows from the video timestep, labeled tau-v here, and action timestep tau-a enter only their corresponding AdaLN blocks. Section 4.1 calls the observation timestep tau-o; these denote the video/observation noise control in the respective presentations. e03e05e07e08

What it supports. The architecture provides a concrete basis for joint modeling: video and action tokens interact inside the network. Algorithm 6 then denoises both streams together. The separate blocks preserve specialist transformations; the figure does not depict a robot controller that merely consumes an already completed video rollout.

Where the evidence stops. The snowflake applies to the upstream VLM, not the trainable understanding expert. This schematic omits the attention-mask specification and detailed control interface; its arrows establish information flow, not measured causal contributions to success.

2. Motivation

2.1 The problem and the proposed response

Source description

Robot policies need language understanding, visual dynamics and executable controls, yet heterogeneous videos lack compatible action labels. Motus addresses both unified prediction and cross-embodiment motion transfer. Its manipulation formulation predicts action chunks from the current image and instruction, optionally proprioception. e02e03

2.2 What this reading follows

A manipulation policy must connect what a scene means, how it could change, and which controls could produce that change. Motus brings those functions into a shared generative network while retaining specialized video, action and understanding branches. Its second bridge is an optical-flow latent action: a compact motion target that makes videos without robot controls useful for action pretraining. Read the architecture and bottleneck first, then the simulation and robot tables. The evidence supports the complete recipe on average, but its partial-success metrics, inconsistent clean ablation means and embodiment-dependent failures require care. This edition reads the supplied December 2025 arXiv v2, including its embedded supplement. e01e03e04e05e11e12e13e14

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryWAMs
ArchitectureDual-system
Prediction paradigmJoint prediction
QuadrantQ3 · Dual-system × Joint prediction

3.1 Evidence-based assessment

Conflicts with the recorded classification

Reader analysis

Joint prediction is supported by shared tri-expert attention and simultaneous video/action denoising. The Dual-system label is questionable: Figure 1 and Section 4.1 describe a unified MoT network with layer-level interaction, not a separately executed world-model/controller cascade. Expert-specific parameters alone do not settle that taxonomy. This reader assessment preserves the recorded classification. e03e08

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • Current visual observation and language instruction; optional proprioception in the policy formulation
  • Noised future video/action tokens and separate flow timesteps; clean future actions or observations when required by the selected mode
  • Future video observations and/or latent-action or embodiment-specific action chunks, depending on training stage and inference mode

4.2 Equations and their role

L=Lrecon+λa∥areal−apred∥2+βLKL\mathcal{L}=\mathcal{L}_{\mathrm{recon}}+\lambda_a\lVert a_{\mathrm{real}}-a_{\mathrm{pred}}\rVert^2+\beta\mathcal{L}_{\mathrm{KL}}
Equation (2) combines flow reconstruction, alignment between real and predicted actions, and latent-space KL regularization. Table 11 sets the alignment weight to 1.0 and KL weight to 10^{-6}. e04e16
ℓθ=ℓactionθ+ℓobsθ,ℓactionθ=E ⁣[∥vaθ−(ϵa−at+1:t+k)∥22]\ell^\theta=\ell^\theta_{\mathrm{action}}+\ell^\theta_{\mathrm{obs}},\qquad\ell^\theta_{\mathrm{action}}=\mathbb{E}\!\left[\left\lVert v_a^\theta-(\epsilon_a-a_{t+1:t+k})\right\rVert_2^2\right]
The action velocity field is trained toward Gaussian noise minus the clean action chunk; observations use an analogous loss. Here t is observation time, k the chunk horizon and theta the learned parameters. The expectation includes examples, noise and the action flow timestep. e02e06e07

5. Method in detail

5.1 Use modality noise to select what the network predicts

Source description

Motus uses the same interacting experts for several different conditional distributions. Begin with a current image and an instruction. For joint prediction, initialize both future video and future action streams as noise, then iteratively update both using their predicted velocity fields. For a VLA, keep the future-video stream noisy throughout while updating actions. For a world model, supply a clean action sequence and update only video; inverse dynamics instead supplies clean future observations and updates actions. VGM mode leaves actions noisy and produces video. Algorithms 2–6 make these differences explicit. Thus the scheduler determines which information is available as a condition, while the MoT architecture supplies cross-modal interaction. A generated future video is not automatically the input to a separate inverse-dynamics controller. e03e06e07e08

Figure 3. Flow reconstruction supplies a compact motion target for action pretraining. Original paper, p. 5 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow the downward reconstruction path. DPFlow first converts image changes into pixel displacements, displayed as a flow image. DC-AE compresses that representation, and the lightweight encoder maps it into the bottleneck labeled z-t. The decoder and lower DC-AE reconstruct flow rather than the original RGB scene. Section 4.2 specifies four 512-dimensional features followed by a 14-dimensional latent. Weak action supervision is described in the text and Equation (2), although its alignment branch is not drawn here. That loss accompanies reconstruction and KL regularization, with labeled examples comprising 10% of the stated mixture. e04e05e16

What it supports. The bottleneck makes motion from videos available as a training target even when robot controls are absent. The intended bridge to controls comes from weak action alignment and later target-robot fine-tuning. Flow reconstruction itself establishes a representation-learning objective; it does not certify motor feasibility.

Where the evidence stops. The diagram draws five feature boxes above and below the bottleneck, whereas Section 4.2 specifies four tokens. The crop is preserved faithfully; the dimensional account follows the text. The source does not resolve this schematic discrepancy.

5.2 Separate motion representation learning from robot specialization

Reader analysis

The latent-action pipeline supplies an intermediate target for videos that have no recorded controls. Optical flow emphasizes displacement, the bottleneck compresses it, and the reconstruction loss checks whether useful motion survives compression. Action alignment on a minority of labeled examples is intended to connect that representation to robotic behavior. Stage 2 then predicts latent actions jointly with video, before Stage 3 learns target-robot controls. Reader interpretation: matching latent dimensionality to a typical action vector is a useful interface choice, but it cannot by itself establish a common motor meaning across embodiments. The Stage 2 ablation supports the complete training intervention, including extra data and compute. To identify the alignment mechanism specifically, a controlled loss ablation is needed, with the rest of that training recipe held fixed. e04e05e12e17

5.3 Match each claimed capability to its evaluation

Reader analysis

The evidence spans different quantities. RoboTwin success measures executed simulated tasks; physical-robot partial success assigns weights to intermediate subgoals; world-model fidelity compares generated observations with robot data; IDM MSE compares predicted and reference controls. These scores cannot be substituted for one another. The dispenser example illustrates the distinction: 96% partial success corresponds to eight complete runs out of ten, not a 96% completion frequency. Reader analysis also changes the generalization claim when the supplement is included: the Agilex OOD cube result favors π0.5 and does not appear in the main platform summary. The strongest defensible conclusion is that the complete Motus recipe improves several aggregate manipulation scores, while mode-specific fairness, uncertainty, latency and some cross-embodiment behavior remain insufficiently resolved. e09e10e13e14e15

5.4 Training and inference

During training

Source description

Latent-action learning mixes 90% unlabeled data with 10% labeled trajectories, including task-agnostic Curobo image/action pairs. Stage 1 adapts only the VGM; Stage 2 trains all three experts on latent actions while freezing the upstream VLM; Stage 3 specializes Motus on target-robot controls. Web priors enter through off-the-shelf models. e04e05

Source description

The model learns two rectified-flow velocity fields from independently noised video/action targets. Training data include Egodex, Agibot, RDT, RoboMind, RoboTwin, task-agnostic and in-house data. Table 13 specifies batch size 256, AdamW and approximately 8,000/10,000/400 GPU-hours across stages. e06e07e17

During inference

Source description

VGM denoises video while keeping future actions noisy; VLA denoises actions while keeping future video noisy. World-model mode fixes clean actions; IDM fixes clean future observations. Joint mode denoises both streams. Algorithms 2–6 specify these modes; the joint-mode prose incorrectly points to Algorithm 2. e07e08

Reader analysis

Table 11 lists ten inference steps. Target-robot training produces control chunks, and Figure 8 depicts execution; video prediction alone does not establish task completion. The source omits measured control latency and a complete feedback/replanning cadence. e05e16e19

5.5 Implementation flow

  1. Fuse complementary experts

    Wan 2.2 5B supplies video priors. Frozen Qwen3-VL-2B features feed a separate understanding expert. Each branch keeps its own Transformer transformations, with shared attention connecting video, action and understanding tokens. Video and action branches receive independent flow timesteps. e03e05

  2. Compress visual motion

    DPFlow produces inter-frame optical flow represented as RGB. Section 4.2 specifies DC-AE compression to four 512-dimensional tokens, followed by a 14-dimensional latent. Reconstruction, action alignment and KL regularization shape this motion target; its dimensionality alone does not guarantee executable control. e04

  3. Predict dense actions with sparse video

    Table 11 lists eight frames at 5 Hz and 48 actions at 30 Hz, reducing redundant video tokens. The RoboTwin IDM experiment instead states a 16-action chunk; these configurations must remain distinct. e03e10e16

6. Experiments & results

Motus connects video, action and understanding experts through shared attention, then uses separate modality noise levels to switch among five prediction modes. Optical-flow latent actions provide pretraining targets for videos without robot action labels. Execution results favor the complete recipe on average, but do not establish uniform generalization or isolate every architectural contribution.

6.1 Read the original evidence

Table 14. Full-task results show high average performance and task-specific exceptions. Original paper, p. 18 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Each method has paired Clean and Rand. columns, so compare the same scene condition across methods. The Average row reports all tasks rather than only the selected rows in the main paper’s Table 2. The evaluation described in Section 5.2 uses multi-task training with clean and randomized demonstrations and 100 execution trials per task. Start with the randomized average, then scan individual tasks to see whether aggregate gains are uniform. In particular, bold entries can belong to baselines or intermediate variants. The last three method groups also permit comparison of no pretraining, video-only pretraining and the full Motus recipe. e09e11e12e17

What it supports. Motus reports 88.66% clean and 87.02% randomized success. X-VLA reports 72.80% and 72.84%, while π0.5 reports 42.98% and 43.84%. Subtracting randomized means gives gains of 14.18 and 43.18 percentage points. The full recipe has the highest displayed average, with individual tasks still favoring alternatives.

Where the evidence stops. Table 2 disagrees with this table on two clean ablation means. Also, the prose’s over-45-point randomized gain over π0.5 is not supported by the average row. Equal fine-tuning steps do not imply equal pretraining compute or data.

Table 3. Physical robot scores measure weighted subgoal progress, with important exceptions. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each platform separately because the task sets differ. These percentages are partial-success scores: completing intermediate subgoals earns credit even if the overall task fails. Section 5.3 specifies 100 training trajectories per task and a joint multi-task model for each platform. The main table summarizes the task outcomes, while Tables 4–5 and 15–16 give the actual subgoal weights and counts. To calibrate the metric, the Agilex dispenser breakdown has eight complete Motus runs and two runs scoring 0.8, producing 96% partial success. Its full-completion frequency is therefore different from the headline partial score. e13e14e15

What it supports. The printed platform means favor Motus: 63.22% versus 14.79% for π0.5 on AC-One, and 59.30% versus 48.60% on Agilex. The result is not uniform. For Agilex oven manipulation, Motus scores 34% and π0.5 scores 36%; the no-pretraining variant also leads on keyboard tasks.

Where the evidence stops. The AC-One cube baseline is 46 here but 57.5 in Table 15. Table 3 omits Agilex cube tasks; Table 16 reports OOD cube scores of 31.25% for Motus versus 68.75% for π0.5. Preserve these limits when citing the averages.

Figure 12. Qualitative joint-mode sequences complement, but do not replace, execution metrics. Original paper, p. 22 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each horizontal strip from left to right; the stacked images inside each strip provide different camera views of the same manipulation setting. The source caption identifies the collection as video-action joint prediction during real-world inference. The examples include coffee-related manipulation, an oven interaction, towel handling and a water dispenser. Use the panels to inspect the depicted progression of arms, objects and contact configurations. For the computational meaning of joint mode, consult Algorithm 6: it updates video and action streams together from noise. These panels do not show the numerical action sequence or mark a per-frame correspondence between predicted and executed controls. e08e10e19

What it supports. The visual shows the types of interactions used to illustrate the joint mode. The quantitative evidence for its control advantage comes separately from Table 8, which reports 87.02% joint versus 83.90% VLA success on randomized RoboTwin. Qualitative robot imagery and that simulation comparison answer different evidentiary questions.

Where the evidence stops. No prediction-versus-execution legend, timestamps or action traces are supplied here. Treat the image as an author-provided qualitative illustration; it cannot establish rollout accuracy, end-to-end completion rates or physical validity of every generated action.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
RoboTwin 2.0 multi-task manipulation

50 tasks; 2,500 clean plus 25,000 randomized demonstrations; 40k fine-tuning steps; 100 execution trials per task.

Motus: 88.66 clean; 87.02 randomized.

Average task success (%)

X-VLA: 72.80/72.84; π0.5: 42.98/43.84 (clean/randomized).

Calculated randomized gains are 14.18 and 43.18 percentage points. The prose claim of over 45 points over π0.5 in randomized scenes conflicts with these means; the clean gain is 45.68 points. Uncertainty is unreported. e09e11

Effect of pretraining stages

RoboTwin 2.0, with downstream training after each pretraining variant.

Randomized: 77.00 without pretraining; 81.86 Stage 1; 87.02 complete recipe.

Average task success (%)

Full versus Stage 1: +5.16 percentage points; versus no pretraining: +10.02.

Supports the combined recipe, without isolating flow supervision, architecture, data diversity or extra compute. Clean means conflict across displays. e11e12

Real-world multi-task manipulation

AC-One and Agilex-Aloha-2; 100 training trajectories per task; one multi-task model per platform.

Table 3: AC-One 63.22; Agilex-Aloha-2 59.30.

Average partial success (%)

π0.5: 14.79/48.60; without pretraining: 25.86/26.60.

Subgoal credit is not binary success. These are printed averages; baseline reporting conflicts and omitted supplementary cube tasks limit their interpretation. e13e14

Get Water from Water Dispenser on Agilex-Aloha-2

Table 5: ten trials per method, with subgoal weights 0, 0.4, 0.8 and 1.0.

Motus: 96% partial score; 8/10 full completions.

Partial success (%) and full completions

π0.5: 62% partial score; 1/10 full completion.

Evidence of executed task progress in a small, task-specific evaluation. e15

Joint versus VLA inference

RoboTwin 2.0 randomized data, Table 8.

Joint 87.02; VLA 83.90.

Average task success (%)

Joint gains 3.12 percentage points, calculated from the reported values.

Favors joint inference in this setting; no latency comparison or uncertainty is reported. e08e10

World-model prediction on robot data

Both real-world platforms; Table 6 average row.

FID 11.209; FVD 61.20865; SSIM 0.8661; LPIPS 0.063645; PSNR 25.0700.

FID/FVD/LPIPS lower; SSIM/PSNR higher

No competing world-model baseline is listed.

Fidelity metrics concern predicted video, not task success. Sample count and detailed split protocol are unspecified. e10e20

Inverse-dynamics action prediction

100 RoboTwin randomized samples; 16-action chunks.

Motus 0.014.

Action MSE (lower)

ResNet18+MLP 0.044; DINOv2+MLP 0.122.

Input parity is unclear: Algorithm 4 uses future observations, but the baseline description specifies current observation. e07e10

LIBERO-Long manipulation

Ten long-horizon tasks; stated standard evaluation protocol.

Motus 97.6.

Average success score

X-VLA 97.6; OpenVLA-OFT 94.5.

A reported tie; fine-tuning and uncertainty details are limited. e18

VLABench manipulation

Three tasks, In Distribution and Cross Category tracks; π0.5 values sourced from its official implementation.

Motus 0.48 in distribution; 0.25 cross category.

Average success fraction

π0.5: 0.43 and 0.22.

Average gains coexist with lower Select Fruit scores in both tracks. e18

Out-of-distribution cube placement

Cube placed outside training space; supplementary Tables 15–16.

AC-One 75; Agilex-Aloha-2 31.25.

Partial success (%)

π0.5: 28.125 on AC-One; 68.75 on Agilex-Aloha-2.

Generalization varies by embodiment; Table 3 omits the Agilex cube results. e14

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Figure 6. Pretraining stages improve the reported means, but the comparison bundles several changes. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Within each scene condition, follow the legend from the lightest no-pretraining bar through Stage 1 to the darker Stage 2 pretraining bar. The latter denotes the complete pretraining recipe evaluated after downstream specialization, not deployment without Stage 3. The left group is randomized; the right group is clean despite the caption emphasizing the randomized setting. Read the numeric labels rather than estimating bar heights: the vertical axis starts near 68 rather than zero. The arrows compare the full recipe with no pretraining. Their percent signs label differences between percentage-valued success rates, so interpret them as percentage-point differences. e05e09e11e12e17

What it supports. Randomized success increases from 77.00% to 81.86% to 87.02%. The final stage adds 5.16 percentage points beyond Stage 1, and the full recipe adds 10.02 points over no pretraining. The plotted clean comparison rises from 77.56% to 88.66%, a difference of 11.10 points.

Where the evidence stops. These clean bars agree with Table 14 but conflict with Table 2’s 72.8/82.86 intermediate means. The ablation changes training stages, data exposure and compute together; it does not isolate optical-flow supervision or shared attention, and no uncertainty is shown.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Figure 6/Table 14 show clean no-pretraining/Stage-1 means 77.56/82.26; Table 2 prints 72.8/82.86. Table 3 gives AC-One cube baseline 46, whereas Table 15 gives 57.5. These conflicts remain unresolved. e09e11e12e13e14

Reader analysis

Improvement across all real-world tasks is not supported: Agilex oven scores are 34 versus 36 for π0.5, and its OOD cube results also favor that baseline. Stage ablations combine several changes and report no seed variability. e12e13e14

Author claim

The authors identify more universal motion priors and learning latent actions from internet-scale general videos as future work. Inherited web priors should not be equated with demonstrated direct latent-action training on arbitrary internet video. e05e21

7.2 Questions for discussion

  1. Would joint inference retain its advantage with matched control latency and identical evaluation seeds? (e08, e10, e16)
  2. How much of Stage 2’s gain comes from action alignment rather than extra heterogeneous data and compute? (e04, e12, e17)

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproduction needs the named foundation models, DPFlow/DC-AE processing, Table 12 datasets and target-robot demonstrations. Table 13 gives GPU-hours but omits GPU model/count and software versions; exact data-mixture weights and the full execution configuration are absent. e03e04e16e17

Reader analysis

Resolve uniform timestep sampling in Section 4.1/Algorithm 1 versus Table 11’s Logit Normal entry, and the 16-action RoboTwin description versus the 48-action configuration. Figure 3 draws five token boxes while its text specifies four; the intended implementation requires clarification. e04e06e07e10e16

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Isolate weak action alignment within Stage 2 pretraining

Reader-proposed experiment: start from the same Stage 1 checkpoint and compare latent-action training with the reported alignment weight 1.0 against weight 0. Keep the 90/10 data mixture, labeled subset, reconstruction/KL terms, number of updates, initialization seeds and Stage 3 demonstrations identical. Evaluate both on identical clean and randomized RoboTwin seeds, reporting per-task success and action-prediction error over repeated training seeds. If alignment is the causal bridge claimed, the aligned variant should improve downstream control beyond the uncertainty of the comparison, rather than merely improve flow reconstruction. Indistinguishable control performance would weaken that mechanism-specific explanation. This check has not been run. e04e05e09e12e16e17

Check 2: Test whether joint inference helps under a matched control budget

Reader-proposed experiment: use one fixed Motus checkpoint and compare Algorithm 5 VLA inference with Algorithm 6 joint inference on identical randomized RoboTwin initial states. First use the same denoising-step count; then compare under equal measured wall-clock inference budgets. Fix the action chunk length, execution interval and feedback cadence, explicitly resolving the source’s 16-versus-48 configuration difference. Record success, latency and action errors across paired seeds. The reported 3.12-point joint advantage should survive repeated trials to support a robust control benefit; disappearance under matched latency would expose a deployment tradeoff. No such latency-matched comparison is reported or reproduced here. e07e08e09e10e16

8.3 Reading coverage

Visual audit: The title/author/version page, all twelve figures, all sixteen tables, rectified-flow equations, six training/inference algorithms and supplementary implementation pages were rendered and visually inspected. All six final original crops were separately viewed at their saved resolution; legends, labels and table rows are retained. Figure 1’s timestep arrows and frozen-VLM marker were checked against Sections 4.1/4.3; Figure 3’s token-count discrepancy, Figure 6’s percentage-point labels and Table 2/Table 14 conflicts are disclosed. All method, numerical and reproduction evidence pages are included above. PDF pp. 10–12 contain references and were read in the complete text pass, but were not visually inspected. No external supplements, code or videos were opened.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22. Appendix coverage: reviewed.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract; Sections 1–3: introduction, related works and problem formulation (PDF pp. 1–4)
  • Sections 4.1–4.3: architecture, latent actions, training and data (PDF pp. 4–6)
  • Sections 5.1–5.4: baselines, simulation, real-world evaluation and ablation (PDF pp. 6–8)
  • Section 6: conclusion and limitations (PDF pp. 8–9)
  • References (PDF pp. 9–12)
  • Embedded Supplementary Material, Sections 7.1–7.2: training algorithms, five inference modes and experiments (PDF pp. 13–14; supplement pp. 1–2)
  • Sections 8.1–8.3 and 9.1–9.3: additional benchmarks, execution results, architecture, datasets and configuration (PDF pp. 15–19; supplement pp. 3–7)
  • Supplementary visualizations (PDF pp. 20–22; supplement pp. 8–10)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Separate supplemental material availability has not been fully verified.
  • The extraction limitation was addressed by inspecting original PDF figures and tables. All seven text chunks were read individually, covering all 22 pages.
  • Identity: PDF p. 1 identifies arXiv:2512.13030v2, dated 25 December 2025. Title and all 16 authors match the catalog. The catalog labels the work CVPR 2026; no venue edition was supplied or compared, and equivalence to that edition is unverified.
  • The embedded ten-page Supplementary Material was read completely. Availability of any separate supplemental files remains unverified.
  • Code, project website and external datasets were not inspected; no experiments were reproduced. Reference-only PDF pp. 10–12 were read as text, not visually inspected.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title/author block and arXiv marginInspect

Observed title is Motus: A Unified Latent Action World Model; all sixteen authors match the catalog. The artifact is arXiv:2512.13030v2, 25 December 2025; affiliations include Tsinghua University, Peking University and Horizon Robotics.

Go to primary source ↓
e02PDF pp. 3–4, Section 3, Embodied Policies, Eq. (1) and Challenges 1–2Inspect

Defines observations, instructions, proprioception, actions and action chunks; motivates unified distributions and learning across heterogeneous action spaces.

Go to primary source ↓
e03PDF p. 2, Figure 1; pp. 4–5, Section 4.1, Model Architecture, Action-Dense Video-Sparse Prediction and Experts DetailsInspect

Tri-model Joint Attention connects separately parameterized video, action and understanding Transformers. Wan 2.2 5B and Qwen3-VL-2B provide pretrained priors. Figure 1 marks the upstream VLM frozen and feeds video/action flow timesteps into their respective AdaLN blocks. Video sampling is sparser than action sampling.

Go to primary source ↓
e04PDF p. 5, Section 4.2, Figure 3 and Eq. (2)Inspect

DPFlow RGB flow is compressed to four 512-dimensional tokens and a 14-dimensional latent. Training combines reconstruction, action alignment and KL with 90% unlabeled and 10% labeled data, including task-agnostic Curobo supervision. Figure 3 depicts five token boxes, unlike the stated four.

Go to primary source ↓
e05PDF pp. 5–6, Section 4.3, Table 1 and Figure 4Inspect

Three stages adapt the VGM, train Motus experts with latent actions and frozen VLM, then fine-tune on target-robot controls. Six data levels distinguish inherited web priors from human, synthetic, task-agnostic, multi-robot and target-robot data.

Go to primary source ↓
e06PDF p. 4, Section 4.1, unnumbered rectified-flow lossesInspect

Action and observation losses regress velocity toward Gaussian noise minus clean future targets, using separate uniformly sampled timesteps; total loss is their sum.

Go to primary source ↓
e07PDF p. 13 (supplement p. 1), Section 7.1, Algorithms 1–4Inspect

Training noises modalities independently. VGM keeps actions noisy, world-model mode conditions on clean actions, and IDM conditions on clean current/future observations. Algorithm 1 uses discrete uniform timesteps in interpolation expressions without explicit normalization details.

Go to primary source ↓
e08PDF p. 14 (supplement p. 2), Section 7.1, Algorithms 5–6Inspect

VLA keeps future observations noisy while denoising actions; joint mode denoises both. The joint-mode prose cites Algorithm 2 although Algorithm 6 gives the joint procedure.

Go to primary source ↓
e09PDF pp. 6–7, Section 5.2 and Table 2, Average rowInspect

The 50-task protocol uses 2,500 clean and 25,000 randomized demonstrations, 40k fine-tuning steps and 100 execution trials per task. Randomization covers backgrounds, clutter, table height and lighting. Table 2 prints clean no-pretraining/Stage-1 means 72.8/82.86; prose claims over 45 points over π0.5 in randomized scenes.

Go to primary source ↓
e10PDF p. 14 (supplement p. 2), Section 7.2, Tables 6–8 and IDM paragraphInspect

Table 6 reports two-platform world-model fidelity without a competitor. Table 7 reports MSE 0.044/0.122/0.014 over 100 samples; baseline text specifies current observations and 16-action chunks. Table 8 reports VLA 83.90 and joint 87.02 success.

Go to primary source ↓
e11PDF p. 18 (supplement p. 6), Table 14, full task rows and Average rowInspect

Motus clean/randomized means are 88.66/87.02, X-VLA 72.8/72.84 and π0.5 42.98/43.84. No-pretraining means are 77.56/77.00 and Stage 1 means 82.26/81.86. Task rows show exceptions to Motus superiority.

Go to primary source ↓
e12PDF p. 8, Section 5.4 and Figure 6Inspect

Randomized bars are 77.00, 81.86, 87.02; clean bars are 77.56, 82.26, 88.66. Arrows show differences 10.02 and 11.10 with percent signs. Variants change pretraining stages; the plot has no error bars.

Go to primary source ↓
e13PDF pp. 7–8, Section 5.3 and Table 3, task and Average rowsInspect

Real-world models train jointly per platform with 100 trajectories per task and use subgoal-weighted partial success. Table 3 gives Motus means 63.22/59.30, π0.5 14.79/48.60, and no-pretraining 25.86/26.60. Agilex oven is 34 versus 36; AC-One cube baseline is printed as 46.

Go to primary source ↓
e14PDF p. 19 (supplement p. 7), Tables 15–16, Grab Cube, OOD, towel and oven rowsInspect

Table 15 prints AC-One cube baseline 57.5. OOD cubes are outside the training space: Motus/π0.5 are 75/28.125 on AC-One and 31.25/68.75 on Agilex. Table 16 contains Agilex cube results absent from Table 3. Towel rows distinguish partial folds from full completion.

Go to primary source ↓
e15PDF p. 8, Table 5, subgoal counts and Partial Success Rate rowInspect

Agilex dispenser weights are 0/0.4/0.8/1.0. Motus has two 0.8 outcomes and eight full successes; π0.5 has five 0.4, four 0.8 and one full success, producing partial scores 96% and 62%.

Go to primary source ↓
e16PDF p. 16 (supplement p. 4), Table 11 and Section 9.1Inspect

Table 11 lists 30-layer action/understanding experts, eight frames at 5 Hz, 48 actions at 30 Hz, ten inference steps, Logit Normal sampling, alignment weight 1.0, KL weight 10^{-6} and total scale 8B. No measured latency or complete feedback policy is provided.

Go to primary source ↓
e17PDF p. 17 (supplement p. 5), Tables 12–13Inspect

Table 12 lists Egodex 230,949; Agibot 728,209; RDT 6,083; RoboMind Franka 9,589 and Aloha 7,272; RoboTwin 27,500; task-agnostic 1,000; in-house 2,000. Stage batches are 256 with AdamW, weight decay 0.01, learning rates 8e-5/5e-5/(1–5)e-5 and approximately 8,000/10,000/400 GPU-hours. Hardware models/counts and software versions are omitted.

Go to primary source ↓
e18PDF pp. 15–16 (supplement pp. 3–4), Section 8.2 and Tables 9–10Inspect

LIBERO-Long has ten tasks; Motus and X-VLA score 97.6, OpenVLA-OFT 94.5. VLABench tests three tasks; average Motus/π0.5 fractions are 0.48/0.43 in distribution and 0.25/0.22 cross category. Select Fruit favors π0.5 in both tracks.

Go to primary source ↓
e19PDF p. 16, Figure 8; p. 22 (supplement p. 10), Figure 12 and captionInspect

Figure 8 is labeled real-world execution demonstrations. Figure 12 presents joint-mode real-world inference as five multi-view sequences, without numerical action traces, per-frame timing or a prediction-versus-execution legend.

Go to primary source ↓
e20PDF p. 15, Figure 7; p. 17, Figure 9; pp. 20–21, Figures 10–11 and captionsInspect

Figures 7 and 9 visualize VGM outputs. Figures 10–11 pair predicted and ground-truth world-model video for Agilex-Aloha-2 and AC-One, supplying qualitative fidelity evidence rather than measured task success.

Go to primary source ↓
e21PDF pp. 8–9, Section 6, Conclusion and LimitationsInspect

Future work includes improved unified architectures, more universal motion priors and latent-action learning from internet-scale general videos.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.