Latent Action as Intention Enables Efficient Future Imagination for World Action Models
1. Paper overview
In one sentence: LAWA replaces test-time future-video generation with continuous latent transition prediction, preserving tested control performance at intermediate inference cost. E3E6E7E9E11E15E17E20
| At a glance | What to know |
|---|---|
| Research problem | Author claim Generating future observations can make world-action policies expensive to deploy. Removing that generation, as in Fast-WAM, is faster but produces weaker generalization in the authors' matched experiments. The paper asks whether an explicit future representation can retain useful predictive conditioning at substantially lower cost than video synthesis. E2E11E17 |
| Core mechanism | |
| A key reported result | RoboCasa full-data manipulation: 80.8% Average task success rate. 24,000 training trajectories across 24 tasks; 50 evaluation trials per task. Matched Fast-WAM: 76.3%; Joint-WAM: 78.8%. LAWA is numerically above both matched alternatives. The table provides no uncertainty establishing statistical superiority or equivalence. E10E11 |
| Reading caution |
Core contributions
Figure 2. The latent branch preserves a future representation after video generation is removed. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the colored branches as different token types: green video, purple latent actions and blue executable actions. Training includes current and future observations, tokenizer-derived targets and demonstrated actions. The attention matrices on the right specify which keys each query can access. Current observation z_0 remains isolated from future information. Latent tokens can use current observations and other latent tokens; action tokens can also use the evolving latent sequence. At inference, future-video tokens disappear but purple and blue tokens are still denoised. The compact matrix therefore exposes the intended deployment interface more directly than the broad label “joint attention” does. E6E7E9E22
What it supports. LAWA retains future-conditioned control without denoising future images at deployment. Its latent and action experts remain distinct modules connected by structured attention. The architecture supports an inference-time latent prediction pathway; it does not support the claim that future prediction is only an auxiliary training task.
Where the evidence stops. The tokenizer’s targets are discrete embeddings, while predicted latent states remain continuous during inference and are not projected back to the codebook. The diagram omits language conditioning; the source text states that it is provided.
2. Motivation
2.1 The problem and the proposed response
Generating future observations can make world-action policies expensive to deploy. Removing that generation, as in Fast-WAM, is faster but produces weaker generalization in the authors' matched experiments. The paper asks whether an explicit future representation can retain useful predictive conditioning at substantially lower cost than video synthesis. E2E11E17
2.2 What this reading follows
LAWA explores a middle option between generating an entire imagined future and discarding future prediction at deployment. It first learns compact transition targets from action-free videos, then trains latent and action experts to denoise together. The future-video branch still contributes during training, but an attention mask prevents the deployable branches from depending on its future tokens. Read the architecture and tokenizer diagrams together: the first explains what runs on the robot, while the second explains where its training targets come from. The result tables then test performance, functional use of the latent sequence, latency and physical execution, with the limits of each comparison kept explicit. E3E6E7E9E11E15E17E20
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 dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Joint prediction |
| Quadrant | Q3 · Dual-system × Joint prediction |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Q3 classification is supported by distinct modality-specific Transformer experts coupled through joint attention, with latent future states and executable actions jointly denoised at deployment. This is architecture evidence beyond joint training alone, and the control path is not a separate inverse-dynamics stage. Joint video-action modeling applies during training. The 'Latent prediction & JEPA' tag is supported through latent prediction, but the stated reconstruction and flow-matching objectives do not establish a JEPA-specific mechanism. Lower measured latency supports the efficiency tag without proving a particular real-time control rate. E6E7E8E9E17
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
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Learn transition targets before learning robot control
LAWA first learns a tokenizer from action-free video. A DINOv2 encoder and a non-causal transformer represent the observation sequence, after which successive contextual feature differences are compressed and assigned to codebook entries. A causal forward decoder must reconstruct the next observation from the previous one and the transition tokens. Mask prediction adds region-level supervision around hands and manipulators, with targets generated by SAM 2. This stage can use both robot and egocentric videos without robot action labels. The source aligns motion speeds through frame sampling and rebalances robot videos to about 20% of expected samples. The tokenizer is then frozen. Its outputs become training targets for the latent expert, while labeled demonstrations teach the separate executable-action expert. E3E4E5E6E8
Figure 3. Action-free video supplies transition targets through reconstruction and mask supervision. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with current and future observations on the left. DINOv2 features are contextualized across space and time, then successive feature differences are compressed and quantized into codebook embeddings. The forward decoder receives the previous observation as well as these transition tokens, making next-frame reconstruction a check on what the tokens retain. The mask output adds hand or manipulator supervision generated with SAM 2. The detailed method uses detached reconstruction features and latent-action prompts for this auxiliary prediction. During action-free pretraining the tokenizer and decoder learn together; during robot-policy training the tokenizer is frozen and supplies latent targets. E3E4E5E8E23
What it supports. The pipeline converts unlabeled robot and egocentric videos into supervision for future transitions. It can train the representation without low-level robot action labels. The mask objective encourages attention to manipulation regions, but its benefit is an empirical design choice rather than a definition of what every latent code means.
Where the evidence stops. The tokenizer uses non-causal sequence context to form training targets. Calling the targets “intentions” is operational: reconstruction and mask prediction do not prove that the codes isolate causal actions, goals or a minimal sufficient representation.
5.2 Keep the future pathway while removing future pixels
Robot-policy training jointly denoises video, latent-action and action targets with independent noise and flow times. The structured attention mask matters as much as the losses: latent and action queries cannot read future-video tokens, so their deployment interface does not depend on those tokens remaining available. At inference, the current observation is encoded once and cached; only latent and action experts are repeatedly evaluated. The latent expert moves continuously through embedding space toward codebook-derived training targets. It does not select discrete codes by nearest-neighbor projection during or after sampling. My interpretation is that the method retains a compact prediction of task transitions while avoiding the cost of rendering their appearance. The information flow supports that interpretation more directly than the word “intention” alone. E6E7E9
5.3 Attribute the benefit to the tested recipe
The main success and latency results should be interpreted together with the egocentric-pretraining comparison. Without that pretraining, LAWA reports 59.7% few-shot and 76.3% full success, below Joint-WAM’s corresponding 63.1% and 78.3%. After pretraining, LAWA reaches 65.6% and 80.8%. Thus compact latent prediction by itself is not uniformly stronger than explicit future-observation prediction in the tested settings. Mask supervision adds a smaller improvement over the no-auxiliary variant, whereas optical-flow supervision reduces performance. My reading is that representation, data and objective choices cooperate in the reported recipe. The latent perturbation experiment shows that the resulting pathway is used; it does not assign the whole improvement to architecture while holding every other factor fixed. E15E16E18
5.4 Training and inference
During training
First train the tokenizer and forward decoder on action-free videos. Per-source frame sampling aligns motion speeds, and weighted sampling keeps robot videos at approximately 20% of expected samples. Reconstruction uses L1 and LPIPS losses; masks use binary cross-entropy, Dice, and IoU losses. Noise-substitution quantization avoids separate codebook losses, and low-usage entries are refreshed. E4E5E8
Then freeze the tokenizer and train the three policy experts with flow matching. Current observations remain clean; each future modality receives independently sampled Gaussian noise and shifted flow time. Padded targets are masked in the losses. E6E9
During inference
Encode the current observation once and cache its features. Omit future-video tokens, initialize latent intentions and action chunks from noise, and jointly denoise their two experts under the training visibility mask. The action expert consumes the evolving continuous latent state; no nearest-neighbor projection occurs during or after denoising. E6E7
The latent sequence conditions executable action prediction directly rather than feeding a separately described inverse-dynamics stage. Although physical trials are reported, the supplied text does not specify how much of each predicted chunk is executed before obtaining feedback and replanning. E6E7E19
5.5 Implementation flow
- Construct transition targets
DINOv2 extracts frame patches. A non-causal transformer contextualizes them with factorized spatial and temporal attention. Differences between successive contextualized features are compressed into L tokens per transition and assigned to nearest codebook embeddings. E3
- Ground the tokenizer
A causal forward decoder reconstructs the next observation from the previous frame and latent transition tokens. A mask decoder receives detached reconstruction features and a projected latent-action prompt, encouraging sensitivity to hands and robot manipulators. E3E4
- Couple modality-specific experts
Video, latent-action, and action experts retain separate Transformer blocks and timestep conditioning while sharing joint attention. Current-observation tokens cannot see future tokens. Latent tokens see current observations and latent tokens; action tokens additionally see the action sequence. Neither control-related branch sees future-video tokens. E7
6. Experiments & results
LAWA retains test-time future prediction through a compact sequence of transition embeddings, jointly denoised with robot actions. A frozen tokenizer supplies discrete training targets, but deployment uses continuous latent states without quantization. Future-video prediction remains a training objective. The complete system improves manipulation success over matched Fast-WAM while reducing latency relative to Joint-WAM; egocentric pre-training is central to this trade-off.
6.1 Read the original evidence
Table 1. The final three rows make the main performance comparison; daggers mark the authors’ implementations. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare methods within the same data column. Few-shot uses 10% of 24,000 training trajectories, corresponding to 100 per task; full uses 1,000 per task. Evaluation covers 24 tasks with 50 trials per task. The matched Fast-WAM and Joint-WAM implementations in the final block are especially informative because they share downstream splits, optimization, batches and training steps with LAWA. Fast-WAM supplies the faster performance baseline, while Joint-WAM retains explicit future-observation generation. Other rows provide context, but many lack few-shot results, and their broader training histories are not controlled by this table. E10E11E14
What it supports. LAWA scores 65.6% few-shot and 80.8% full, compared with Fast-WAM’s 56.0% and 76.3%. The gains are 9.6 and 4.5 percentage points. Joint-WAM reaches 64.1% and 78.8%, making the comparison consistent with retaining its performance level through a less expensive future representation.
Where the evidence stops. The matched downstream recipe does not match parameter counts, modality branches or pretraining objectives. No seed variability or confidence interval appears in the table, so the smaller margins over Joint-WAM should not be treated as established statistical superiority.
Table 5. Compact future prediction occupies the middle of the measured speed tradeoff. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The unit is milliseconds per end-to-end action-chunk prediction, measured on one NVIDIA A800. Read lower values as faster inference, and pair this table with the success comparison rather than treating latency as an isolated objective. Fast-WAM removes test-time future prediction and is fastest. Joint-WAM predicts future observations and is slowest. LAWA predicts latent intentions plus actions, avoiding future-video denoising while retaining an extra branch beyond Fast-WAM. The important comparison is therefore two-sided: how much success LAWA retains relative to Joint-WAM, and how much inference time it adds relative to Fast-WAM. E14E17
What it supports. LAWA requires 338.5 ms versus Joint-WAM’s 593.1 ms, a reported reduction of 42.9%. It remains slower than Fast-WAM at 196.5 ms. Together with Table 1, the measurements support an intermediate performance-latency tradeoff, not the fastest action generator among the tested variants.
Where the evidence stops. The supplied main text defers full latency details to an unavailable appendix. Chunk prediction latency alone does not give the executed control frequency, and batching, denoising-step counts, feedback cadence and timing variability remain unresolved.
Table 7. The proposed system’s advantage extends to the four tested physical tasks. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the table in three pairs of rows, one pair for each demonstration budget. At full data, each of Gear, Battery, Block and Laboratory has 200 demonstrations; 25% therefore means 50 per task. Evaluation uses 20 trials per task and policy. Gear and Battery emphasize assembly, while Block and Laboratory require multistage manipulation. Comparing within a pair holds the stated robot demonstration proportion fixed. Comparing LAWA at 25% with Fast-WAM at 100% asks a different question about the complete recipe’s sample efficiency. The Avg column summarizes task successes, not visual-prediction accuracy or partial subgoal credit. E19E20
What it supports. At 25% data, LAWA averages 40.0% success versus Fast-WAM’s 8.8%; at full data the values are 67.5% and 33.8%. LAWA’s 50-demonstration result exceeds the full-data Fast-WAM average. These measurements concern physical execution on the xArm7, strengthening the evidence beyond simulation.
Where the evidence stops. Twenty trials per task leave substantial sampling uncertainty, and only one platform with four tasks is tested. The comparison evaluates the complete training recipe, including action-free pretraining, rather than isolating the latent architecture under equal total data exposure.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboCasa few-shot manipulation 24 tasks; 100 demonstrations per task, corresponding to 10% of training trajectories; 50 evaluation trials per task. | 65.6% Average task success rate | Matched Fast-WAM: 56.0%; Joint-WAM: 64.1%; DIAL: 58.3%. The improvement over matched Fast-WAM is 9.6 percentage points. This is simulated task completion, not a video prediction metric. E10E11 |
| RoboCasa full-data manipulation 24,000 training trajectories across 24 tasks; 50 evaluation trials per task. | 80.8% Average task success rate | Matched Fast-WAM: 76.3%; Joint-WAM: 78.8%. LAWA is numerically above both matched alternatives. The table provides no uncertainty establishing statistical superiority or equivalence. E10E11 |
| LIBERO-Plus zero-shot robustness Train on original LIBERO trajectories only; evaluate perturbations across four suites without LIBERO-Plus fine-tuning. | 74.4% Micro-averaged success rate | Matched Fast-WAM: 60.0%; Joint-WAM: 70.4%; OpenVLA-OFT: 69.6%. The 14.4-point gain over matched Fast-WAM supports improved robustness under this transfer protocol. Ambiguous perturbation-specific cells are excluded. E12E13 |
| Inference-time latent pathway intervention Full-data RoboCasa; perturb only the model-visible latent-action state. | Unperturbed: 80.8%; isotropic Gaussian noise with σ = 1.0: 52.2%; temporal shuffle: 56.4%. Success rate | Both interventions reduce success relative to the intact latent sequence. The executor uses latent content and temporal organization. These destructive interventions do not establish that the representation has uniquely identifiable intention semantics. E15 |
| Egocentric pre-training comparison RoboCasa; matched video exposure with paradigm-specific objectives and trainable modules. | LAWA without ego: 59.7% / 76.3%; with ego: 65.6% / 80.8%. Few-shot / full-data success rate | Fast-WAM gains 1.5 / 1.7 points; Joint-WAM gains 1.0 / 0.5 points. Without ego, Joint-WAM reaches 63.1% / 78.3%. LAWA benefits most under these schedules, but latent prediction without egocentric pre-training trails explicit future-video prediction. E14E16 |
| Tokenizer auxiliary-supervision ablation RoboCasa with latent actions and egocentric pre-training enabled. | No auxiliary loss: 64.8% / 79.3%; flow: 63.5% / 78.6%; mask: 65.6% / 80.8%. Few-shot / full-data success rate | Mask supervision improves the no-auxiliary variant by 0.8 / 1.5 points; optical flow lowers success. The tested mask objective helps downstream control modestly. This comparison does not directly measure the semantic content of the learned codes. E18 |
| Action-chunk inference efficiency End-to-end prediction on one NVIDIA A800 GPU. | 338.5 ms Latency per action chunk; lower is better | Joint-WAM: 593.1 ms; Fast-WAM: 196.5 ms. The reported reduction against Joint-WAM is 42.9%. LAWA remains slower than Fast-WAM; chunk latency alone does not establish a robot control frequency. E14E17 |
| Physical assembly and long-horizon manipulation xArm7; Gear, Battery, Block, and Laboratory tasks; 200 demonstrations per task at full data; 20 evaluation trials per task. | LAWA: 40.0% at 25% data; 67.5% at full data. Average executed-task success rate | Fast-WAM: 8.8% at 25% data; 33.8% at full data. With 50 demonstrations per task, LAWA exceeds full-data Fast-WAM's reported average. These are physical execution results, limited to the four tested tasks and reported trial counts. E19E20 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. The action expert uses both the latent content and its temporal arrangement. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. This intervention table changes the model-visible latent-action state at inference while leaving the learned policy in place. The unperturbed column is the reference. Gaussian corruption uses isotropic noise with sigma = 1.0, disrupting latent content; temporal shuffling disrupts sequence order. Read the success drop as a test of functional dependence, rather than as another training ablation. It asks whether action generation would behave similarly if the latent pathway carried altered information. Both perturbed columns show substantial degradation. Their comparison does not rank the two interventions by inherent severity because their distributions and affected structure differ. E15
What it supports. RoboCasa full-data success falls from 80.8% to 52.2% under noise and to 56.4% under temporal shuffling. The executor therefore does not simply ignore latent predictions. The results support a useful, temporally organized conditioning pathway, without proving that its representation is uniquely optimal.
Where the evidence stops. Large corruptions can damage any used feature stream. Without norm-matched controls or alternative latent representations, the drops do not independently establish interpretable intentions, causal disentanglement or robustness to modest prediction errors.
7. Analysis & limitations
7.1 What the evidence leaves open
Matched downstream splits and optimization do not equal matched capacity or compute: branches, objectives, parameter counts, and inference costs differ. The results establish the complete recipe's performance more directly than a representation-only advantage. E14E16
Tokenizer targets use non-causal sequence context. Their interpretation as intentions is operational; neither perturbation sensitivity nor mask supervision proves a minimal causal representation of manipulable dynamics. E3E6E15E18
Success tables lack reported confidence intervals or seed variability. Physical evaluation covers one robot platform and four tasks, so broad deployment reliability remains unestablished. E11E13E19E20
Important implementation and data details are deferred to an unavailable appendix, preventing a complete audit of pre-training fairness and inference configuration. E5E14
7.2 Questions for discussion
- Would temporally aligned ground-truth tokenizer targets improve control over predicted intentions, separating latent-prediction error from action-expert error?
- Does LAWA retain its advantage when parameter counts and pre-training compute are matched across paradigms?
- How do latent horizon and denoising steps affect success, feedback frequency, and physical execution latency?
8. Reproducibility audit
8.1 Requirements and known gaps
Target the supplied arXiv v2 artifact. Reproduction requires the video backbone, tokenizer configuration, action-free corpora and sampling recipe, SAM 2 mask-generation procedure, and labeled downstream trajectories. E1E3E4E5E10
Recover codebook size, token dimensions, loss coefficients, optimizer settings, flow-time schedules, denoising steps, action horizon, and execution/replanning policy before comparing performance or latency. These are not fully specified in the supplied main text. E3E8E9E14E17
A focused mechanism test would repeat the latent-content and temporal-shuffle interventions using identical evaluation initializations and multiple training seeds, reporting uncertainty alongside success. E15
The abstract promises a future code and model release. No code was supplied or inspected, and no experiments were reproduced in this review. E21
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: Control the severity of latent interventions
Reader-proposed: repeat temporal shuffling and Gaussian corruption with fixed initial states and action-sampling noise. Add norm-preserving latent permutations, small-noise sweeps and an equally sized corruption of a non-latent conditioning stream. Report success and action divergence over shared seeds. A disproportionately large dependence on temporal latent order would strengthen the transition-sequence interpretation beyond generic sensitivity to damaged features. E7E15
Check 2: Match data and compute in the efficiency comparison
Reader-proposed: train the three paradigms with the same action-free clips, downstream splits and measured accelerator budget, then sweep denoising steps to produce success-versus-latency curves on one device. Record parameter counts and action-feedback cadence. If LAWA’s advantage persists on a shared compute frontier, it would support the compact future interface independently of unequal optimization cost; otherwise the improvement should remain attributed to the complete recipe. E14E16E17E18
8.3 Reading coverage
Visual audit: Original PDF pages 3, 5, 6 and 7 were visually inspected. Final 200-DPI crops reproduce Figures 2 and 3 plus Tables 1, 3, 5 and 7. Every final crop was re-opened after extraction, including the architecture legend and both attention matrices. Other visible tables were read with the relevant source text for context. Appendix material referenced by the main text was not supplied; code and experiments were not inspected or run.
PDF pages inspected for this edition: 3, 5, 6, 7. Appendix coverage: not read.
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
- Title, authors, and arXiv version footer
- Abstract
- 1 Introduction
- 2 Related Work
- 3 Method
- 3.1 Latent Action Tokenizer
- 3.2 Latent Action as Intention
- 3.3 Training Objective
- 4 Experiments
- 4.1 Experimental Setup
- 4.2 Performance on Simulation Benchmarks
- 4.3 Comparison between Paradigms
- 4.4 Ablation Study
- 4.5 Performance on Real-World Tasks
- 5 Conclusion
- References
Outside the original text pass
- Figure images were unavailable and were not visually inspected.
- Some perturbation-specific cells in Table 2 are concatenated; uncertain column assignments and values are omitted.
- The main text repeatedly references an appendix that was not supplied. Separate supplemental material availability remains unverified.
- 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 visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
E1PDF p. 1, title, author block, and arXiv footer
The title and listed authors match the catalog identity; the footer identifies arXiv:2608.24882v2, dated 1 September 2026.
Go to primary source ↓E2PDF p. 1, Abstract and Section 1
The authors motivate compact future intentions through the latency of future-observation denoising and weaker generalization of their matched Fast-WAM implementation.
Go to primary source ↓E3PDF pp. 3–4, Section 3.1, Architecture, Eq. (1)
DINOv2 features receive non-causal spatial/temporal contextualization; successive feature differences are compressed and quantized. A causal forward decoder reconstructs the next observation.
Go to primary source ↓E4PDF p. 4, Section 3.1, Manipulation-centric Auxiliary Supervision
A SAM-style decoder uses detached reconstruction features and latent-action prompts to predict hand/manipulator masks generated with SAM 2.
Go to primary source ↓E5PDF p. 4, Section 3.1, Action-free Egocentric Pre-training
Robot and egocentric clips train the tokenizer and forward decoder. Frame sampling aligns motion speeds; weighted sampling assigns approximately 20% expected samples to robot videos. Further details are deferred to the appendix.
Go to primary source ↓E6PDF p. 4, Section 3.2, opening paragraphs
The frozen tokenizer supplies discrete codebook embeddings. The latent expert predicts continuous states jointly with actions, without nearest-neighbor projection at inference.
Go to primary source ↓E7PDF p. 4, Section 3.2, Multi-model Joint Attention
Experts retain modality-specific blocks and timestep conditioning. The asymmetric visibility mask excludes future-video information from latent/action branches; inference caches current features and evaluates only latent/action experts.
Go to primary source ↓E8PDF p. 4, Section 3.3, Eq. (2) and following paragraph
Tokenizer training combines L1, LPIPS, and mask losses; mask loss includes binary cross-entropy, Dice, and IoU. Noise-substitution quantization and refreshing low-usage entries manage the codebook.
Go to primary source ↓E9PDF p. 4, Section 3.3, flow construction and Eqs. (3)–(4)
Independent modality noise and shifted flow times define velocity targets. Weighted video, latent, and action flow losses train the policy, excluding padded targets.
Go to primary source ↓E10PDF pp. 4–5, Section 4.1, RoboCasa
The protocol uses 24 tasks, 24,000 full-data trajectories, a 10% few-shot subset, and 50 evaluation trials per task.
Go to primary source ↓E11PDF p. 5, Table 1, DIAL and matched WAM rows, Few-shot/Full columns
LAWA reports 65.6/80.8, Fast-WAM 56.0/76.3, Joint-WAM 64.1/78.8, and DIAL 58.3/70.2 percent success. No uncertainty columns are provided.
Go to primary source ↓E12PDF p. 5, Section 4.1, LIBERO-Plus
Models train only on original LIBERO trajectories; evaluation reports perturbation performance across four suites and a micro-average without augmented-data fine-tuning.
Go to primary source ↓E13PDF p. 6, Table 2, Total column; PDF p. 5, Section 4.2, Zero-shot Transfer Results
Overall success is 74.4% for LAWA, 60.0% for matched Fast-WAM, 70.4% for Joint-WAM, and 69.6% for OpenVLA-OFT.
Go to primary source ↓E14PDF p. 5, Section 4.1, Implementation Details
Matched implementations share downstream splits, optimization, batch sizes, and training steps, but differ in branches, objectives, parameter counts, and inference costs. Full latency details are deferred.
Go to primary source ↓E15PDF p. 6, Table 3; PDF p. 5, Section 4.3, The Effectiveness of Latent Action
Perturbing only the visible latent state changes success from 80.8% to 52.2% with isotropic Gaussian noise at σ=1.0, or 56.4% with temporal shuffling.
Go to primary source ↓E16PDF p. 6, Table 4 and Section 4.3, The Effectiveness of Egocentric Pre-training
Without/with ego, LAWA reports 59.7/65.6 few-shot and 76.3/80.8 full; Fast-WAM 54.5/56.0 and 74.6/76.3; Joint-WAM 63.1/64.1 and 78.3/78.8. Video exposure is matched, objectives and modules differ.
Go to primary source ↓E17PDF p. 6, Table 5 and Section 4.3, Performance–latency Trade-off
One A800 yields action-chunk latencies of 196.5 ms for Fast-WAM, 593.1 ms for Joint-WAM, and 338.5 ms for LAWA; the text reports a 42.9% reduction against Joint-WAM.
Go to primary source ↓E18PDF p. 7, Table 6 and Section 4.4
With latent actions and ego pre-training, no auxiliary loss gives 64.8/79.3, flow 63.5/78.6, and mask 65.6/80.8 percent few-shot/full success. The authors caution that mask benefits do not prove representation semantics.
Go to primary source ↓E19PDF p. 7, Section 4.5, Setup and tasks
Physical tests use an xArm7 with a gripper, a RealSense D435 base camera, and two fisheye wrist cameras. Four tasks have 200 demonstrations each and 20 evaluation trials per policy and task.
Go to primary source ↓E20PDF p. 7, Table 7, 25% and 100% data rows, Avg. column; Section 4.5, Results
At 25% data, LAWA/Fast-WAM average 40.0/8.8 percent success; at full data they average 67.5/33.8. The text identifies 25% data as 50 trajectories per task.
Go to primary source ↓E21PDF p. 1, Abstract, final sentence
Code and model release is stated as future work; the sentence does not establish present availability.
Go to primary source ↓E22PDF p. 3, Figure 2, training/inference attention matrices and component legend (visually inspected)
The original architecture diagram distinguishes video, latent-action and action tokens. Its inference matrix removes future-video slots while retaining current observation, latent-action and action tokens; action queries can attend to latent actions. The language instruction is omitted from the drawing.
Go to primary source ↓E23PDF p. 3, Figure 3; p. 4, Section 3.1 (visually inspected diagram and primary text)
The action-free pretraining diagram routes observation sequences through DINO, attention and quantization into a forward decoder with observation and mask outputs. The text describes sequence-contextual feature differences, discrete targets, detached reconstruction features for mask prediction and SAM 2 target masks.
Go to primary source ↓8.5 Primary sources
Latent Action as Intention Enables Efficient Future Imagination for World Action Models ↗
PDF · 4,963 extracted words
Source fingerprint
c9bbd43cf12f09a02f88202e18e929de804d603b57621f4c51298b684c3d52e9