PAPER REPORTENAll readings ↗

LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

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

Authors: Jialei Chen; Kai Wang; Kang Chen; Shuaihang Chen; Feng Gao; Wenhao Tang; Zhiyuan Li; Weilin Liu; Zhuyu Yao; Boxun Li; Yuanbo Xu; Chao Yu

Affiliations: Tsinghua University; Jilin University; Nankai University; Peking University; Harbin Institute of Technology; Zhongguancun Academy; Striding.AI; Infinigence AI

Source: 2606.15768 ↗ · Project page ↗ · Catalog record

Reading: 161 / 558 · 6 original figures & tables · ~19 min ·

1. Paper overview

In one sentence: LaWAM makes predicted DINO features an explicit action-generation input, exchanging pixel-level future synthesis for a compact dynamics interface whose evidence is strongest in stable-camera manipulation. problemarchitecturefactorizationliberoreal-worldlimitationsreproduction-gaps

At a glanceWhat to know
Research problem
Source description

A current-observation VLA lacks an explicit predicted scene change, while pixel-space world-action models spend time synthesizing appearance. LaWAM asks whether a spatial future feature can supply useful dynamics to action generation at lower latency. Its target is chunk-level control, not photorealistic video or an independently validated physical simulator. problemfactorization

Core mechanism
Source description

Retain the latent-action decoder as a 230M-parameter world model, making its predicted future an explicit policy input during deployment. architecture

A key reported resultLIBERO success and action-chunk latency: 98.6% average; Long 97.0, Goal 98.4, Object 99.6, Spatial 99.4; 187 ms.

Success rate (%) and mean model-only latency (ms/chunk). Four suites, 40 tasks, 50 trials/task (2,000 total). Model-only timing: 1,000 predictions on an A100, 10 denoising steps.

Cosmos-Policy: 98.5%, 1413 ms. LingBot-VA: 98.5%, 4482 ms. Only a 0.1 percentage-point average success margin, without reported uncertainty; the larger difference is latency. The approximately 24× ratio is relative to LingBot-VA model inference, not robot task duration. liberolibero-protocollatency-protocol

Reading caution
Source description

Camera shake or large viewpoint changes can dominate the transition and disrupt latent-action learning. Sparse fine cloth dynamics and feature resolution also limit subtle deformation prediction. limitations

Core contributions

  • Source description

    Retain the latent-action decoder as a 230M-parameter world model, making its predicted future an explicit policy input during deployment. architecture

  • Source description

    Combine posterior-to-policy latent-action distillation, insulated dynamics conditioning and native-frequency action tokens aligned by elapsed physical time. stage-twotime-alignment

Figure 2. The decoder survives pretraining and becomes the policy's future-feature interface. Original paper, p. 2 ↗

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

How to read it. Read the upper panel from the two observed frames through frozen DINO and the inverse-dynamics encoder to the latent action. The current-feature bypass also enters LaWM: a transition variable alone does not specify a scene. In the lower panel, latent-action queries replace the future-dependent encoder as the source of the decoder's input. The dotted double-headed distillation line denotes a training match, not bidirectional deployment. Follow the subgoal through Knowledge Insulation and the junction labeled concatenation into the expert; the plus-shaped symbol is not evidence of elementwise addition. A separate VLM branch supplies semantic context, while noise enters the action-denoising pathway below. architecturefactorizationstage-onestage-twoimplementationlatency-protocol

What it supports. The retained LaWM decoder predicts a subgoal in one pass, but action generation still uses iterative denoising. Equation (3) separates the latent prior, deterministic future-feature decoder and action expert. This architecture supports the dual-system interpretation because the future representation is an explicit input to a distinct action generator.

Where the evidence stops. The diagram reuses F_T for future features; the text distinguishes observed target u_T from policy prediction hat u_T. Knowledge Insulation blocks action-expert gradients, not every training update. The future-observation teacher is unavailable at deployment.

2. Motivation

2.1 The problem and the proposed response

Source description

A current-observation VLA lacks an explicit predicted scene change, while pixel-space world-action models spend time synthesizing appearance. LaWAM asks whether a spatial future feature can supply useful dynamics to action generation at lower latency. Its target is chunk-level control, not photorealistic video or an independently validated physical simulator. problemfactorization

2.2 What this reading follows

Imagine giving a robot a spatial hint about where its arm and scene should be after the next action chunk. LaWAM learns that hint in a frozen visual feature space, then lets a separate action expert turn it into executable motion. The key decision is to retain the forward decoder of a latent-action model and use it during deployment. This reading follows the training-to-inference handoff, the success-versus-latency evidence, and the diagnostics behind the dynamics claim. The paper reports strong simulated and physical manipulation results, while leaving camera-motion robustness, fine cloth dynamics and several reproduction details unresolved. problemarchitecturefactorizationliberoreal-worldlimitationsreproduction-gaps

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 paradigmIDM
QuadrantQ4 · Dual-system × IDM

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The recorded Dual-system × IDM classification is supported by an explicit LaWM-to-action-expert boundary and the future-conditioned action factorization. Stage-one latent inverse dynamics supplies teacher transitions; deployment uses a learned prior followed by forward prediction and a distinct action expert with an inverse-dynamics role. Joint training does not make this a single joint future/action generator, and LaWM is used at inference rather than only as an auxiliary loss. architecturefactorizationimplementation

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 primary-view RGB observation and task instruction; optional auxiliary RGB views for policy context
  • Training only: horizon observations, robot end-effector labels and language-annotated trajectories for policy integration
  • Predicted latent action and horizon DINO feature map
  • Executable end-effector action chunk from the flow-matching expert

4.2 Equations and their role

p(a1:T,u^T,z^o,l)=pθ(z^o,l)pω(u^Tu,z^)pη(a1:To,l,u,u^T)p(a_{1:T},\hat u_T,\hat z\mid o,l)=p_\theta(\hat z\mid o,l)\,p_\omega(\hat u_T\mid u,\hat z)\,p_\eta(a_{1:T}\mid o,l,u,\hat u_T)
Equation (3): o is the current observation, l the instruction, u its frozen visual feature, and a_{1:T} the action chunk. The policy prior predicts latent action hat z; LaWM deterministically produces future feature hat u_T; the action expert generates actions. factorization
LLAM=u~TuT22+g(s,z)sT22+βDKL ⁣(qϕ(zu,uT)N(0,I))\mathcal L_{\mathrm{LAM}}=\|\tilde u_T-u_T\|_2^2+\|g(s,z)-s_T\|_2^2+\beta D_{\mathrm{KL}}\!\left(q_\phi(z\mid u,u_T)\,\|\,\mathcal N(0,I)\right)
Equation (4): u_T is the observed horizon feature and tilde u_T its decoder prediction. The posterior q_phi infers z from the transition. Auxiliary predictor g maps current end-effector state s and z to horizon state s_T; beta weights regularization. stage-one
Hb=round(τhb),tb,i=i/hbH_b=\operatorname{round}(\tau h_b),\qquad t_{b,i}=i/h_b
Appendix C.3: branch b runs at frequency h_b; H_b counts action tokens over duration tau. Token i receives a physical timestamp t_{b,i}, preventing equal token indices at different frequencies from being treated as equal elapsed times. time-alignment

5. Method in detail

5.1 Learn the transition with hindsight, then remove hindsight

Source description

The first stage has privileged access to two observations: what the scene looks like now and what it looks like at the selected horizon. Frozen DINO features make the supervision a feature-prediction problem. The posterior compresses that observed change into a continuous latent action, and the decoder learns to recover the future feature using the current scene as context. The auxiliary state head pushes the latent action toward embodied motion, while KL regularization organizes its distribution. Deployment cannot invoke this posterior because the future has not happened. Stage two therefore teaches latent-action queries to predict the posterior's output from present context and instruction. The retained decoder becomes useful precisely at this handoff: it translates a predicted transition into a scene-grounded visual subgoal that the action expert can consume. stage-onefactorizationimplementationstage-two

5.2 Keep future prediction distinct from executable motion

Source description

The model's latent action is an input to LaWM, not an end-effector command. LaWM expands it into a spatial feature map, which joins current features in the dynamics stream of Alternate-DiT. The expert also receives semantic VLM context and produces an action chunk through conditional flow matching. Thus the single-pass efficiency claim applies to future-feature prediction; the action generator still denoises. The policy uses RGB observations without proprioceptive inputs, even though stage-one auxiliary supervision can involve state labels. Knowledge Insulation protects the subgoal pathway from action-expert gradients while the explicit feature loss still supervises it. The source's heatmaps provide a way to inspect this division of labor: they compare a selected arm patch with predicted future patches and overlay that signal on actual observations during execution. factorizationstage-twoimplementationlatency-protocolvisualization

5.3 Separate three different meanings of dynamics evidence

Reader analysis

Reader analysis: the paper provides evidence at three levels that answer different questions. The component ablation asks whether the policy benefits from the complete future-conditioning interface. The feature-rollout probe asks whether predicted representations evolve toward future observations instead of staying at the initial state. The simulated and physical success tables ask whether executable actions complete tasks in the evaluated environments. None subsumes the others: a high DINO similarity need not imply a correct grasp, and high success does not identify which module caused it. Shared latent-action rollouts across screenshots add a qualitative transfer argument, but they do not execute actions in those target embodiments. The reported camera-motion and cloth-resolution limitations further suggest that this interface's usefulness depends on which scene changes the frozen feature representation and training mixture can preserve. ablationrolloutliberorobotwinreal-worldcross-embodimentlimitations

5.4 Training and inference

During training

Source description

Stage one uses roughly 3,000 robot-video hours and 1,500 human-video hours. Future-feature error and KL regularization train the latent model; an auxiliary end-effector predictor encourages embodied motion and is later discarded. Training runs 100k steps on 16 H100s, batch 1024, AdamW learning rate 3×10^-4, weight decay 10^-2 and KL weight 10^-5. stage-onestage-twotraining-config

Source description

Stage two uses language-annotated robot trajectories only: 200k steps on 64 H100s, batch 1024. Distillation and subgoal losses each receive weight 0.1 alongside action flow matching. Learning rates are 10^-4 for the expert and 3×10^-5 for other modules. Knowledge Insulation blocks action-expert gradients from overwriting LaWM; it does not imply LaWM receives no subgoal supervision. stage-twotraining-config

Source description

Benchmark post-training follows integration: LIBERO uses 25k steps/batch 256 after failed-demonstration removal; RoboTwin uses 100k steps/batch 1024 with 2,500 clean and 25,000 randomized demonstrations. Policies receive 256×256 RGB images without proprioception; stage-one auxiliary state supervision is a separate training pathway. implementationlibero-protocolrobotwin-protocol

During inference

Source description

Predict the latent action from present visual-language context, decode one future feature in a single LaWM pass, then denoise the action chunk using both context and subgoal. The default evaluation uses 10 action-denoising steps. Predicted latent actions are not robot commands; the expert supplies executable actions. The described policy has no candidate-rollout search or reward-based planner. factorizationstage-twolatency-protocol

Reader analysis

Execution visualizations hold a subgoal across a chunk and show later chunks receiving new subgoals. Their heatmaps locate similarity to an initial arm patch; they visualize feature-space guidance rather than pixel prediction or a calibrated contact model. visualization

5.5 Implementation flow

  1. Learn a visual transition variable

    Frozen DINOv3 ViT-B/16 encodes current and future observations. A 24-layer posterior transformer infers a continuous latent action. A separate 24-layer decoder predicts future features from current features and that action, conditioning through adaptive layer normalization. stage-oneimplementation

  2. Replace access to the future

    Stage two trains Qwen3-VL latent-action queries to predict the posterior's teacher action using present observation and instruction. The first 16 backbone layers feed a four-block Alternate-DiT expert, which alternates semantic context with current/predicted-future features. With the specified causal ordering, latent-action queries precede auxiliary views; action queries follow them. factorizationimplementation

  3. Align the action horizon

    Keep native control frequencies and add sinusoidal physical timestamps to action queries. Pad and mask variable-length chunks. The aligned robot horizon is 1.2 seconds in LaWM training; human-video transitions use a separate 0.4-second horizon, so fixed duration should not be read as universal across the entire mixture. time-alignmenttraining-config

6. Experiments & results

LaWAM turns a latent-action model's forward decoder into an inference-time source of visual subgoals. A language-conditioned policy predicts a latent transition, LaWM expands it into DINO future features, and a separate action expert generates a robot action chunk. The main appeal is near-leading manipulation success with inexpensive future prediction; the evidence remains concentrated on stable-camera manipulation.

6.1 Read the original evidence

Table 1. The strongest efficiency evidence pairs nearly tied success with a large model-only latency reduction. Original paper, p. 6 ↗

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

How to read it. Start with the rightmost Average column, then move left to check whether a method wins every suite. LaWAM's 98.6% average is only just above the 98.5% entries for Cosmos-Policy and LingBot-VA; it does not lead Long or Object. Now compare the latency column: 187 ms for LaWAM, 1413 ms for Cosmos-Policy and 4482 ms for LingBot-VA. The timing protocol uses an A100 and ordinarily ten denoising steps, averaging one thousand predictions. Blank latency entries mean no value is given. The task results cover forty LIBERO tasks with fifty trials each, rather than forty independent training runs. liberolibero-protocollatency-protocol

What it supports. The roughly 24× latency ratio comes from comparing 4482 with 187 ms. The table therefore supports competitive manipulation performance with substantially cheaper inference than that comparator. It does not support a similarly large increase in task success or a 24× reduction in physical execution time.

Where the evidence stops. No uncertainty intervals establish the significance of the 0.1 percentage-point success gap. Baseline provenance mixes original-paper and reproduced values. Timing excludes execution overhead, and WAM size accounting excludes video-diffusion VAEs and text encoders.

Table 3. Three physical tasks extend the evidence beyond simulated manipulation. Original paper, p. 7 ↗

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

How to read it. Read each task column separately before comparing the average. Pick-and-place and drawer opening use a Franka Panda, whereas towel folding uses a Quanta X1 bimanual robot. The policy receives one hundred fifty demonstrations for each Franka task and two hundred eighty for folding. Each reported percentage summarizes thirty trials, with initial configurations fixed across the compared methods. The source pools seen and unseen conditions, so this table cannot isolate either group's success. LaWAM leads each listed column; π0.5 is its closest average comparator. The zero in LingBot-VA's folding column is a measured result for this setup, not a statement about all cloth tasks. real-worldreal-protocollimitations

What it supports. LaWAM's 90.0% average exceeds π0.5's 83.3% here. Its folding result is particularly relevant to timely control because the towel can move while a policy computes. Unlike a latent rollout illustration, these entries concern actions executed on physical robots and evaluated for task success.

Where the evidence stops. The authors associate slow inference with LingBot-VA's folding failures, but this table does not isolate latency from architecture and training. Thirty trials per task and pooled seen/unseen conditions limit claims about generalization and statistical superiority.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
LIBERO success and action-chunk latency

Four suites, 40 tasks, 50 trials/task (2,000 total). Model-only timing: 1,000 predictions on an A100, 10 denoising steps.

98.6% average; Long 97.0, Goal 98.4, Object 99.6, Spatial 99.4; 187 ms.

Success rate (%) and mean model-only latency (ms/chunk)

Cosmos-Policy: 98.5%, 1413 ms. LingBot-VA: 98.5%, 4482 ms.

Only a 0.1 percentage-point average success margin, without reported uncertainty; the larger difference is latency. The approximately 24× ratio is relative to LingBot-VA model inference, not robot task duration. liberolibero-protocollatency-protocol

RoboTwin clean and randomized bimanual manipulation

50 tasks; 100 trials per task in each setting after mixed clean/randomized training.

Clean 92.64; randomized 89.80.

Average success rate (%)

Fast-WAM 91.98/90.52; LingBot-VA 91.50/90.92, clean/randomized respectively.

Best listed clean average, but below both comparators under randomization. The abstract's 91.22% equals the mean of LaWAM's two setting averages; retain the separate protocols. Table 4 exposes weak tasks, including Open Microwave at 41/43%. identityrobotwinrobotwin-protocol

Physical pick-and-place, drawer opening and towel folding

Franka Panda for the first two tasks; Quanta X1 for folding. Training: 150 demonstrations per Franka task, 280 for folding. Thirty fixed-comparison trials per task include seen/unseen conditions.

93.3 / 86.7 / 90.0; average 90.0.

Success rate (%)

π0.5: 86.7 / 80.0 / 83.3; average 83.3.

Executed robot success supports transfer within these three setups. Seen/unseen results are pooled, and no uncertainty intervals establish statistical superiority. real-worldreal-protocol

LIBERO component ablations

Figure 6 compares removal of LaWM, distillation, KI plus distillation, and pretraining.

Removing LaWM gives the largest degradation, especially on Long.

Success-rate ordering; exact bar values not supplied

Removing KI plus distillation performs below removing distillation alone.

Supports the subgoal interface, but combined removal is not an independent KI-only ablation. Full-model bars differ visually from Table 1; the source does not reconcile the configurations. ablation

Mixed-frequency training and latent rollout diagnostics

From-scratch LIBERO 5/10/20 Hz co-training from shared demonstrations; separate open-loop probe over 500 LIBERO trajectories.

Time encoding recovers performance toward the 20 Hz-only reference; rollout features track future features while diverging from initial features.

Success-rate trend and feature cosine similarity

Mixed-frequency training without encoding degrades; Figure 10 compares rollout/future, future/initial and rollout/initial similarities.

These support temporal alignment and nontrivial feature evolution. They do not establish calibrated physical dynamics or executed cross-embodiment transfer. mixed-frequencyrolloutcross-embodiment

6.3 Ablations and diagnostic examples

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

Figure 6. The ablation supports the complete subgoal interface while leaving some causal effects entangled. Original paper, p. 8 ↗

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

How to read it. Use the legend to compare the five variants within each suite, especially Long and Average. Blue is complete LaWAM, orange removes pretraining, green removes distillation, purple removes both Knowledge Insulation and distillation, and cyan removes LaWM. The cyan bar's largest separation on Long motivates the claim that explicit subgoal conditioning matters for extended tasks. Green versus purple examines the additional loss of insulation when distillation is already absent. That comparison differs from testing insulation while keeping distillation present. The vertical axis begins at 90%, so visual bar-height differences should not be read as proportional differences in total success. ablationstage-twolibero

What it supports. The plotted ordering favors the complete architecture, and removal of LaWM is most damaging. Distillation also helps the policy drive the pretrained decoder. These observations support the proposed interface, but the design does not provide a complete factorial test of whether distillation and gradient insulation have independent or interacting effects.

Where the evidence stops. Exact values and uncertainty are not printed. Figure 6's full-model bars visibly differ from Table 1's reported full-model values; the source supplies no reconciliation. Treat the plot as a separate comparison, not a numerical decomposition of the 98.6% headline.

Figure 7. Elapsed-time encoding repairs much of the ambiguity introduced by mixed control frequencies. Original paper, p. 16 ↗

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

How to read it. Compare cyan and blue within each task suite: both use mixed-frequency demonstrations, but only blue includes the physical-time encoding. Orange is the native 20 Hz-only reference. Appendix C.4 constructs all branches from the same original demonstrations by downsampling to 10 and 5 Hz, then trains from scratch. This holds the source tasks, embodiment, visual domain and instructions fixed; the lower-frequency branches are not new demonstrations. Appendix C.3 explains the mechanism: action-token timestamps are measured in seconds and added through sinusoidal encodings, while the future-feature target is paired to a physical horizon. Equal token indices can therefore denote different elapsed times. time-alignmentmixed-frequency

What it supports. Encoding restores much of the performance lost when frequencies are mixed, with the clearest visual recovery on Long and Goal. This controlled experiment makes a stronger case for temporal alignment than an uncontrolled comparison between different robot datasets would, although the native-rate reference still performs better.

Where the evidence stops. The source reports unlabeled bars without uncertainty. Recovery toward the 20 Hz reference does not establish that adding real heterogeneous datasets always helps; this construction changes temporal sampling without adding tasks, embodiments or independent trajectories.

Figure 10. The aggregate feature probe tests whether LaWM predicts change instead of copying the initial state. Original paper, p. 19 ↗

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

How to read it. Read the blue curve together with the green and dashed gray curves. Blue compares predicted rollout features with ground-truth future features. Green compares the rollout with the initial features, while gray compares the true future with the initial features. High blue similarity alone could be misleading if scenes barely changed. The declining initial-state similarities show why this control matters: the predictions remain closer to the future as both future and prediction move away from the starting features. The source averages this probe over five hundred LIBERO trajectories. The horizontal label is Step, so the plot itself should not be converted into seconds. rolloutcross-embodiment

What it supports. The combination of high future agreement and departure from the initial features supports nontrivial latent evolution. It is evidence against a simple copy-the-current-feature explanation for this aggregate probe. It complements, rather than replaces, the executed-task results and qualitative examples of context-specific latent transitions.

Where the evidence stops. Cosine similarity in DINO space is not a calibrated measure of contact dynamics or control reliability. No uncertainty bands appear. Appendix D.4 does not fully specify the latent-action sourcing protocol for this aggregate probe, limiting deployment-level interpretation.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Camera shake or large viewpoint changes can dominate the transition and disrupt latent-action learning. Sparse fine cloth dynamics and feature resolution also limit subtle deformation prediction. limitations

Reader analysis

The broad tables mix reported and re-evaluated baselines rather than controlling pretraining data and compute. Small success margins, qualitative transfer rollouts and unreported uncertainty limit stronger causal/generalization claims. Parameter accounting excludes WAM video VAEs and text encoders. latency-protocolrobotwinreal-worldcross-embodiment

7.2 Questions for discussion

  1. Would replacing predicted future features with capacity-matched current features preserve performance?
  2. Can latency-matched cloth trials separate the benefit of better subgoals from the benefit of fresher observations?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Reproduction needs the DINO/Qwen architecture, exact data mixture, teacher posterior, gradient boundaries, physical timestamps and end-effector label conversions. RoboTwin's reported post-training alone takes about 20 hours on 64 H100s; low inference latency does not imply inexpensive training. implementationstage-twotime-alignmentrobotwin-protocol

Open question

Missing details include latent dimensionality/query counts, dataset sampling weights, precise end-effector conventions, missing-state handling for human-video auxiliary supervision, software versions and seeds. Resolve these before claiming an exact reproduction. reproduction-gaps

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: Factor distillation and gradient insulation independently

Reader-proposed check, not performed: train a matched 2×2 grid with latent-action distillation on/off and Knowledge Insulation on/off, retaining LaWM in every arm. Fix the pretrained checkpoint, demonstrations, update count, loss weights for unchanged terms and evaluation initial states; repeat across seeds. Log teacher–prior latent error, predicted-future feature error and success by LIBERO suite. The paper's mechanism predicts that removing insulation damages feature fidelity and that distillation improves prior alignment. If either changes training losses without changing task success, or their effects depend strongly on each other, the independent-contribution interpretation would need revision. This fills the missing KI-only comparison rather than assigning an exact effect from Figure 6. stage-twoablationlibero-protocoltraining-config

Check 2: Control latency while measuring towel-folding success

Reader-proposed check, not performed: keep a trained LaWAM policy fixed and insert controlled delays between observation acquisition and action dispatch on the same towel-folding initial configurations. Include its ordinary timing as the control, and log observation age, actual dispatch latency, failures and task success across repeated trials. If available implementations permit, compare other policies at matched observation ages as a second arm, without changing denoising settings mid-comparison. A success decline under added delay would support the source's stale-observation mechanism. A persistent LaWAM advantage at matched delay would suggest additional representational or training benefits. A null delay effect would weaken attributing the reported folding gap chiefly to latency. real-worldreal-protocollimitationslatency-protocol

8.3 Reading coverage

Visual audit: Visually inspected the title, twelve authors, eight affiliations and v1 date on page 1; all Figures 1–15 and Tables 1–4; method equations on pages 3–5; architecture, physical-time encoding, hardware and optimization on pages 14–16; evaluation protocols and robot setups on pages 16–18; and rollout/qualitative appendices through page 23. All six final original crops were separately inspected with legends, axes or table headers intact. Architecture arrows, the distillation line, frozen-DINO markers and Knowledge Insulation/concatenation labels were checked against Sections 3.1–3.3 and Appendix C.2. Figure 6's discrepancy with Table 1 is retained and disclosed. Reference pages 9–12 were read as text only. No external supplement, linked media, code or cited work was inspected; no reproduction experiment was run.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23. 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
  • 1 Introduction
  • 2 Related Work
  • 3 Method
  • 3.1 Problem Formulation
  • 3.2 Latent World Model
  • 3.3 Latent World Action Model
  • 4 Experiments
  • 4.1 LIBERO
  • 4.2 RoboTwin
  • 4.3 Real-World Experiments
  • 4.4 LaWM Dynamics Analysis
  • 4.5 Component Ablations
  • 5 Limitations
  • 6 Conclusion
  • References
  • A Detailed RoboTwin Results
  • B Qualitative Visualization Protocol
  • C Architecture and Implementation Details
  • C.1 Latent World Model Architecture
  • C.2 Latent World Action Model Architecture
  • C.3 Physical-Time Alignment for Mixed-Frequency Data
  • C.4 Mixed-Frequency Training Experiment
  • C.5 Training Details
  • D Experimental Details
  • D.1 LIBERO Protocol
  • D.2 RoboTwin Protocol
  • D.3 Real-World Protocol
  • D.4 Additional LaWM Rollout Analysis

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.
  • Identity/version: the inspected title page identifies arXiv:2606.15768v1, 14 June 2026; the title and all twelve authors match the supplied paper identity. All eight printed affiliations are retained in the verified metadata. No other edition was supplied or compared.
  • The supplied extraction does not reconstruct figure images; this gap was resolved by inspecting the retained PDF, including every figure and table. Reference pages 9–12 were read as text, without a separate visual pass.
  • Separate supplemental material availability has not been fully verified; no separate supplement was supplied.
  • Code, linked websites and cited papers were not inspected. No experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

identityPDF p. 1, title, author/affiliation block, arXiv margin stamp and AbstractInspect

The title page reads LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies and lists twelve authors, in the order retained in the verified metadata. The artifact is arXiv:2606.15768v1, dated 14 June 2026. Eight affiliations are printed, including Striding.AI and Infinigence AI. The abstract reports a RoboTwin headline success rate of 91.22%.

Go to primary source ↓
problemPDF pp. 1–3, Introduction, Figure 1 and Related WorkInspect

The authors motivate feature-space future conditioning by the cost of iterative pixel generation and distinguish their retained forward decoder from latent-action methods that discard it.

Go to primary source ↓
architecturePDF p. 2, Figure 2 and caption; pp. 4–5, Sections 3.2–3.3Inspect

The two-stage diagram routes frozen DINO features through a transition posterior and forward decoder, then uses policy-predicted latent actions to obtain a subgoal for an Alternate-DiT action expert. Knowledge Insulation protects the dynamics interface from action-expert gradients.

Go to primary source ↓
factorizationPDF pp. 3–4, Section 3.1, Eqs. (1)–(3)Inspect

The policy prior predicts a latent action from observation and instruction; LaWM deterministically predicts a future feature; the separate action expert conditions on current context and that feature. The training posterior requires an observed future and cannot serve as the deployment prior.

Go to primary source ↓
stage-onePDF p. 4, Section 3.2, Eq. (4)Inspect

LaWM training combines future-feature squared error, an auxiliary end-effector horizon-state predictor and KL regularization toward a standard normal. The auxiliary head is discarded, while the posterior supplies stage-two teacher actions.

Go to primary source ↓
stage-twoPDF p. 5, Section 3.3, Eq. (5), and Section 4Inspect

Policy integration combines latent-action distillation, predicted-subgoal supervision and conditional action flow matching. LaWM pretraining uses roughly 3,000 robot-video hours and 1,500 egocentric-human-video hours; policy integration uses language-annotated robot trajectories only.

Go to primary source ↓
implementationPDF p. 14, Appendices C.1–C.2Inspect

DINOv3 ViT-B/16 features feed 24-layer posterior and decoder transformers, with adaptive layer normalization for latent-action conditioning. The policy uses Qwen3-VL's first 16 layers, four Alternate-DiT blocks totaling 16 layers, hidden size 1024, causal query ordering, 256×256 RGB images and end-effector action labels; it receives no proprioceptive inputs.

Go to primary source ↓
time-alignmentPDF pp. 14–15, Appendix C.3, Eqs. (6)–(7)Inspect

Native-frequency branches use a common elapsed horizon within the aligned mixture, with rounded horizon-times-frequency token counts. Action tokens receive sinusoidal encodings of timestamps in seconds; variable-length batches use padding and masks.

Go to primary source ↓
mixed-frequencyPDF pp. 15–16, Appendix C.4 and Figure 7Inspect

The controlled from-scratch experiment downsamples the same 20 Hz LIBERO demonstrations to 10 and 5 Hz. Physical-time encoding improves mixed-frequency performance toward the 20 Hz-only reference without adding expert trajectories. The figure has no exact bar labels or uncertainty intervals.

Go to primary source ↓
training-configPDF pp. 15–16, Appendix C.5Inspect

LaWM uses 16 H100 GPUs, 100k steps, AdamW at 3×10^-4, weight decay 10^-2, batch 1024 and KL weight 10^-5. Horizons are 1.2 s for robot teleoperation and 0.4 s for human videos. Policy integration uses 64 H100 GPUs, 200k steps, batch 1024, loss weights 0.1 each, and learning rates 10^-4 for the action expert and 3×10^-5 for other modules.

Go to primary source ↓
latency-protocolPDF p. 6, Table 1 caption; p. 16, Appendix C.5, evaluation paragraphInspect

Latency excludes simulator/robot execution overhead. The stated measurement averages 1,000 action-chunk predictions on an A100, ordinarily with 10 denoising steps. WAM parameter counts exclude video-diffusion VAEs and text encoders. Baseline values mix original-paper and reproduced results.

Go to primary source ↓
liberoPDF p. 6, Table 1, LaWAM, Cosmos-Policy and LingBot-VA rowsInspect

LaWAM reports 2.3B parameters, 187 ms latency and success rates 97.0/98.4/99.6/99.4 for Long/Goal/Object/Spatial, averaging 98.6%. Cosmos-Policy and LingBot-VA each average 98.5%; their latencies are 1413 and 4482 ms, respectively.

Go to primary source ↓
libero-protocolPDF p. 16, Appendix D.1; p. 6, Table 1 captionInspect

LIBERO training removes failed demonstrations, runs 25k steps with batch 256, and evaluates 40 tasks over 2,000 trials, or 50 trials per task.

Go to primary source ↓
robotwinPDF p. 6, Table 2 caption and Average row; p. 13, Table 4, Average, Hanging Mug and Open Microwave rowsInspect

LaWAM clean/randomized averages are 92.64/89.80%, versus Fast-WAM 91.98/90.52% and LingBot-VA 91.50/90.92%. LaWAM scores 51/43% on Hanging Mug and 41/43% on Open Microwave. Fast-WAM and LingBot-VA are re-evaluated from open weights; other baseline results are taken from cited papers.

Go to primary source ↓
robotwin-protocolPDF p. 16, Appendix D.2Inspect

RoboTwin trains on 2,500 clean and 25,000 randomized demonstrations, uses 100k steps and batch 1024, and reports 100 evaluation trials per task for each setting. Training takes about 20 hours on 64 H100 GPUs.

Go to primary source ↓
real-worldPDF p. 7, Table 3 and Section 4.3Inspect

Over 30 trials per task, LaWAM reports 93.3% pick-and-place, 86.7% drawer opening and 90.0% towel folding, averaging 90.0%. The π0.5 row is 86.7/80.0/83.3%, averaging 83.3%; LingBot-VA reports 0.0% towel folding.

Go to primary source ↓
real-protocolPDF pp. 17–18, Appendix D.3 and Figure 9Inspect

Franka Panda tasks use 150 demonstrations each; Quanta X1 towel folding uses 280. Initial configurations are fixed across methods, with seen and unseen conditions. Drawer trials cover 10 configurations with rotational perturbations up to 15 degrees. Folding includes flattening, two long-edge folds and one short-edge fold.

Go to primary source ↓
ablationPDF p. 8, Section 4.5 and Figure 6; p. 6, Table 1Inspect

Figure 6 compares complete LaWAM with removal of pretraining, distillation, KI plus distillation, or LaWM. Removing LaWM produces the largest degradation, particularly on Long. The plot has no exact bar labels or uncertainty; its full-model bars visibly differ from the main Table 1 values, without a reconciliation in Section 4.5.

Go to primary source ↓
limitationsPDF p. 8, Section 5; p. 18, Appendix D.3, Failure inspectionInspect

The authors identify camera-motion-dominated transitions and scarce fine cloth dynamics as limitations. They also describe insufficient feature resolution for subtle deformations and associate LingBot-VA towel failures with stale observations during slow inference.

Go to primary source ↓
visualizationPDF p. 5, Figure 3; pp. 13–14, Appendix B; p. 17, Figure 8; pp. 20–21, Figures 11–13Inspect

Subgoal heatmaps compare a selected initial robot-arm patch with predicted DINO patches by cosine similarity. Observations and feature visualizations show execution approaching predicted arm regions; they are not reconstructed future RGB frames.

Go to primary source ↓
rolloutPDF p. 19, Figure 10 and Appendix D.4Inspect

Across 500 LIBERO trajectories, predicted-future versus true-future feature similarity remains high while predicted and true futures become less similar to initial features. The chart gives aggregate cosine-similarity curves without uncertainty or a detailed latent-action sourcing protocol for this aggregate probe.

Go to primary source ↓
cross-embodimentPDF p. 7, Figure 5 and caption; p. 8, Section 4.4; pp. 22–23, Figures 14–15Inspect

A source video's inferred latent-action sequence is applied from initial observations in other scenes/embodiments, producing context-specific feature rollouts. Some target contexts are unseen screenshots. These are latent rollout demonstrations rather than executed transfer trials.

Go to primary source ↓
reproduction-gapsPDF pp. 14–18, Appendices C.1–C.5 and D.1–D.3Inspect

The implementation description supplies architectures, optimizers, hardware and benchmark schedules, but does not fully specify latent-action dimensionality/query counts, dataset sampling proportions, end-effector label conventions, treatment of missing auxiliary state labels in human data, software versions or random seeds.

Go to primary source ↓

8.5 Primary sources

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