DELE-w0.5: Inferring Action from Future Latent State for Robotic Manipulation
1. Paper overview
In one sentence: DELE-w0.5 trains actions alongside compact future-state latents but removes future-observation tokens at deployment, gaining a shorter inference sequence while leaving the causal benefit of future-state supervision unablated. E02E03E04E05E06E07E09E10E12E14
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that dense future-video reconstruction spends capacity on appearance and intermediate transitions that manipulation does not require. They propose predicting a compact endpoint representation alongside actions, aiming to retain useful physical-outcome supervision while reducing visual generation overhead. E02 |
| Core mechanism | |
| A key reported result | Overall physical task completion: 62.5% (50/80), reported from real-robot execution Full-task success. Astribot S1; four fixed-scene tasks; 20 trials per method-task pair; 180-second budget; no physical assistance. The complete comparison contains 640 trials, including 80 for DELE-w0.5. XR0 is the strongest baseline on this metric: 30.0% (24/80). The difference is 32.5 percentage points. This measures satisfaction of complete physical task conditions, not visual prediction quality. No confidence intervals are reported. E07E08E09 |
| Reading caution | Reader analysis The central future-to-action explanation conflicts with the explicit action-to-future attention direction and action-only deployment. Equation (19)'s arguments and the noise-stream self-attention description also leave the exact implementation of visibility insufficiently resolved. The report follows the explicit Section 4.4 rules without treating them as code-verified behavior. E02E05E06 |
Core contributions
- Source description
An asymmetric attention policy allows future tokens to access actions while preventing the reverse connection; future tokens are removed during inference. E06
Figure 2. The training model predicts two targets from clean conditioning and noisy inputs. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read upward from the four inputs. Language and the current image enter the condition stream on the left; their encoders and projections form a common representation. The action chunk and encoded next observation enter the noise stream on the right, where training injects noise. Both streams produce query, key, and value representations through timestep-dependent modulation before the masked attention block and prediction heads. The condition timestep is fixed at τ = 1, while the target-stream timestep is sampled during training. Although the diagram labels a next-state image at the input, the supervised prediction lives in the vision encoder’s latent space. The following mask figure is needed to determine which target can actually read another. E03E04E05E06iv01
What it supports. The architecture replaces future RGB reconstruction with a latent future-state target and jointly trains action prediction. Its compact representation can share learning with control, while the clean condition stream retains language and current visual evidence. The diagram alone does not establish a sequential future-generation-then-action pipeline; token visibility determines the operational dependency.
Where the evidence stops. Figure 2 is a training overview. Section 4.4 removes future-observation groups at deployment, and the paper’s broad future-state narrative should not be used to add an unshown inference-time planning or inverse-dynamics stage.
2. Motivation
2.1 The problem and the proposed response
The authors argue that dense future-video reconstruction spends capacity on appearance and intermediate transitions that manipulation does not require. They propose predicting a compact endpoint representation alongside actions, aiming to retain useful physical-outcome supervision while reducing visual generation overhead. E02
2.2 What this reading follows
DELE-w0.5 proposes that robot control should learn action-relevant state change without reconstructing full future video. Its training diagram joins language and current-image conditioning with noisy action and future-latent targets. The crucial detail appears in the attention mask: actions cannot read predicted future observations, and those future tokens are removed at inference. This makes the deployed mechanism more specific than the title’s suggestion of inferring actions from a generated future. The illustrated reading follows that distinction, then examines four real-robot tasks through complete success, ordered-stage progress, and timing. The reported gains are substantial in this suite, but neither baseline comparisons nor failure heatmaps isolate the future-state objective’s causal contribution. E02E03E04E05E06E07E09E10E12E14
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 | One Model |
| Prediction paradigm | Joint prediction |
| Quadrant | Q1 · One Model × Joint prediction |
3.1 Evidence-based assessment
Insufficient evidence to decide
One Model has architectural support: the streams feed concatenated QKV tensors into a common masked-attention prediction block, rather than a separately described world model and inverse-dynamics controller. Joint prediction describes the two training targets, but deployment predicts actions alone. The complete snapshot therefore needs qualification: Joint video-action modeling overstates endpoint-latent prediction, and latent prediction is documented without establishing a JEPA-specific objective. Q1 is defensible if it includes training-time auxiliary future prediction; it should not imply inference-time future-conditioned action generation. E03E04E05E06
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 Compress the visual target, then learn a velocity field
The policy takes language, current multi-view images, and proprioception and predicts a sixty-step action chunk; the dual-arm action space has fourteen dimensions per step. Qwen3 encodes language and DINO-v3 encodes images, with projections into a shared 1024-dimensional hidden space. During training, a future image is encoded by the vision encoder rather than reconstructed as RGB. Actions and future latents are interpolated with Gaussian noise: A^τ = τA + (1 − τ)ε, with an analogous expression for the future latent. The model regresses the velocity toward the clean target, and the overall objective is L = Lₐ + λLₒ. This defines a joint representation-learning problem for motor commands and visual state change. It does not, by itself, establish that a future-state sample is computed before choosing an action at deployment. E03E04E05
Figure 3. The action-to-future attention cell is masked, and the future group disappears at inference. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Rows are queries and columns are keys or values. In the training panel, the language and current-observation rows read only those two conditioning groups. The action row is colored under language, current observation, and action, but white under future observation. The future-observation row can read all four groups. This direction matters: a future query may read the action; an action query may not read the future. On the right, inference retains the same conditioning/action submatrix after deleting the future group. Multi-view images are combined into each observation label in this schematic, so the size of a colored block does not count individual image or action tokens. E05E06iv02
What it supports. Under the explicitly described mask, action generation uses present conditioning and relationships within the action chunk. Future-state prediction can influence learned shared parameters through training, but it is not a visible input to the action query. Removing its token group therefore preserves the stated action visibility pattern at deployment.
Where the evidence stops. Equation 19 still lists a noisy future latent as an argument of the action velocity function. That notation is ambiguous relative to Figure 3 and Section 4.4; an implementation-level mask audit remains necessary to resolve it fully.
5.2 Use the mask to distinguish supervision from an online dependency
The title and discussion frame action-relevant futures as an intermediate representation, but the explicit visibility rules narrow that claim. Figure 3 and Section 4.4 allow future-observation queries to attend to actions, while action queries cannot attend to the future-observation group. At inference, future groups are removed entirely. My interpretation is that future prediction can shape shared parameters during training while the deployed action path remains conditioned on current information. This is compatible with auxiliary predictive learning, but does not demonstrate inference-time action selection by searching or inverting a generated future. Equation 19 complicates the account because its action velocity function still lists the noisy future latent as an argument. Resolving whether that is shorthand, redundant notation, or a different implemented dependency requires checking the actual masks across every attention block. E02E05E06
5.3 Check terminal conditions before accepting a progress gain
The evaluation’s ordered stages make incomplete behavior visible, but the terminal physical state remains the success criterion. A Door trial requires an opening beyond forty-five degrees sustained for three seconds; Pepsi requires retaining the can while closing the refrigerator. A trial that advances far and then misses that condition can score high normalized progress while still failing. Because each task has a different number of stages, the paper averages s/K within each task and then macro-averages tasks. Table 2 and Figure 6 should therefore be read together: one summarizes performance and the other localizes losses along the ordered sequence. My deduction is that DELE’s advantage concerns more reliable transitions through this fixed task suite, not simply detecting the first object interaction. The qualitative intervention examples remain separate early observations, without a counted generalization benchmark. E07E08E09E12E13
5.4 Training and inference
During training
Training combines squared-error action and future-latent velocity losses with weight λ. The target future representation comes from the vision encoder rather than a pixel or intermediate-video reconstruction objective. E03E05
The comparison uses common task fine-tuning data. Baselines follow their official configurations for three equivalent epochs; this does not establish matched pretraining or compute. Encoder freezing, the proposed model's detailed training schedule, and the numerical loss weight are not specified. E04E05E07E14
During inference
Future-observation groups are removed while context/action visibility remains unchanged. The model produces actions without generating future visual tokens. E06
This deployment description does not contain a predicted-goal-to-action inverse-dynamics stage. It also does not specify the flow solver, sampling-step count, executed fraction of each chunk, or replanning schedule; removing visual tokens alone does not establish single-step action sampling. E05E06
5.5 Implementation flow
- Encode the current context
Qwen3 encodes language and DINO-v3 encodes images. Linear projections map their features to a common hidden width of 1024. Concatenated language and current-image tokens pass through a shared self-attention block. E03E04
- Construct conditioning attention features
The condition stream applies RMS normalization and learned timestep-dependent scaling before QKV projection, with L2 normalization of queries and keys. Its timestep embedding is fixed at τ = 1. E04
- Corrupt both prediction targets
The noise stream receives interpolations between Gaussian noise and clean action chunks or encoded future observations. A uniformly sampled τ sets the interpolation level. The paper describes analogous attention processing to obtain noise-stream QKV features. E05
- Predict through coupled attention
Condition- and noise-stream QKV tensors are concatenated for masked attention and output projection. Under Section 4.4's visibility rules, context reads context, actions read context and actions, and future tokens read every group. The described direction therefore lets future prediction depend on action information. E05E06
6. Experiments & results
DELE-w0.5 trains a manipulation policy with action-chunk generation and future visual-latent prediction, then removes future-observation tokens at deployment. A central tension is that the stated attention mask prevents actions from accessing predicted futures, despite the paper's future-to-action narrative. Its concrete deployment description therefore supports action generation with auxiliary future-state training rather than explicit inference from a generated future. The policy completes 50 of 80 physical trials across four fixed-scene tasks; the experiments do not isolate the contribution of future prediction.
6.1 Read the original evidence
Table 2. Complete success and partial progress tell related but different stories. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each pair of task columns separates normalized ordered-stage progress from complete physical success. Progress averages s/K, where s is the highest consecutively completed stage and K is the task’s stage count; it does not count arbitrary later actions after a missed prerequisite. Success requires the full terminal condition. Each method receives twenty trials on each of four tasks, so Overall Succ. covers eighty trials. Macro Prog. weights the four task-level progress means equally despite different stage counts. Read the strongest baseline separately for each metric: GWP0.5 leads baseline macro progress, whereas XR0 leads baseline overall success. This prevents mixing two distinct reference points when reporting improvements. E07E08E09E14iv03
What it supports. DELE-w0.5 reports 81.3% macro progress and 62.5% complete success, corresponding to 50 successful trials out of 80. XR0 achieves the highest baseline complete success at 30.0%. Add Ice remains DELE’s hardest task by full completion at 45.0%, even though its normalized progress is 63.3%.
Where the evidence stops. The comparison uses fixed scenes, one robot, shared task data, equivalent three-epoch adaptation, and predefined resets. It does not isolate the future-state loss from architecture or pretraining differences, or establish broad out-of-distribution generalization.
Figure 5. Inference time, successful execution time, and probability of completion must remain separate. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Panel (a) places macro stage progress on the horizontal axis and complete-task success on the vertical axis. Panel (b) keeps overall success but changes the horizontal axis to median core-model inference latency on an RTX 4090; moving left means a faster call, not necessarily a faster completed task. Panels (c) through (f) instead use measured robot completion time and condition that time on successful trials. Read each task separately, since its sequence and success probability differ. Method labels sometimes use abbreviations inherited from Table 2. DELE is highlighted in red throughout, but the competing methods occupy different tradeoffs rather than a single ordering by speed. E09E10E11iv04
What it supports. DELE combines the highest measured success in this suite with relatively low core-model latency. Its successful Door executions appear much shorter than its successful Pepsi and Add Ice executions, illustrating that one model’s inference speed does not determine all task durations. The lower plots preserve success probability alongside the conditional timing statistic.
Where the evidence stops. Core latency excludes network transfer, image decoding, action projection, IK, trajectory processing, and robot execution. Completion-time centers omit failed trials; they cannot be interpreted as expected time to success across all attempted rollouts.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Overall physical task completion Astribot S1; four fixed-scene tasks; 20 trials per method-task pair; 180-second budget; no physical assistance. The complete comparison contains 640 trials, including 80 for DELE-w0.5. | 62.5% (50/80), reported from real-robot execution Full-task success | XR0 is the strongest baseline on this metric: 30.0% (24/80). The difference is 32.5 percentage points. This measures satisfaction of complete physical task conditions, not visual prediction quality. No confidence intervals are reported. E07E08E09 |
| Ordered manipulation progress Each trial scores its consecutively completed stages divided by the task's stage count; task means are then equally weighted across four tasks. | 81.3%, reported Macro normalized ordered-stage progress | GWP0.5 obtains 61.3%. The paper reports a 20.1-percentage-point improvement using unrounded values. Progress credits valid task prefixes and is distinct from complete success. Add Ice consolidates eight raw annotation stages into six semantic stages. E08E09 |
| Door opening; Pepsi retrieval; Add Ice; microwave popcorn 20 real-robot trials per task under the common protocol | Door: 80.0% (16/20); Pepsi: 65.0% (13/20); Add Ice: 45.0% (9/20); microwave: 60.0% (12/20) Task-specific full success | Best baseline successes are respectively 45.0% (GWP0.5, XR0, and π0.5 tied), 30.0% (XR0), 25.0% (XR0), and 20.0% (XR0). DELE-w0.5 leads each task. Add Ice remains its least reliable task, and the stage analysis places remaining difficulty around acquisition, transfer, release, and terminal conditions. E09E12 |
| Policy computation Core-model inference on an NVIDIA RTX 4090 | 87.5 ms, reported measurement Median inference latency | The prose reports faster inference than GWP0.5, π0.5, LingBot-VLA2, and HY-VLA, and comparable latency to Spirit-v1.5. Exact baseline plot coordinates are omitted. Timing excludes network transfer, image decoding, action projection, inverse kinematics, trajectory processing, and robot execution. It is not end-to-end control latency. E10 |
| Successful task execution duration DELE-w0.5's successful trials only: Door n = 16, Pepsi n = 13, Add Ice n = 9, microwave n = 12 | Door: 16.66 s; Pepsi: 45.92 s; Add Ice: 46.27 s; microwave: 39.50 s Mean measured completion time | The paper states that methods share a runtime measurement protocol; numerical baseline means are not recoverable from the supplied prose. These are physical execution times conditional on success. They exclude failed trials and should not be interpreted as expected time to accomplish a task. E11 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 6. The failure diagnostic localizes where partial progress stops becoming success. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read a method row left to right within one task block. Each cell estimates P(s ≥ k), the fraction of twenty trials reaching stage k while completing all earlier stages. Because these are ordered prefixes, values cannot increase as the task advances. Full is the final physical success condition, not merely another attempted action. Compare the Door and Microwave blocks first: many methods contact or open the appliance successfully, then lose trials during coordinated execution or terminal-state completion. Use Table 1’s stage definitions to interpret the numbered columns. The common color scale allows cross-panel scanning, but different tasks have different numbers and meanings of intermediate stages. E08E12E14iv05
What it supports. DELE reaches Door’s third stage in every trial but finishes in 80%, locating its failures at sustained opening. In Pepsi, its reach drops from 90% after extraction to 65% at handoff and remains 65% at Full. Add Ice falls from 95% at its first stage to 45% completion, revealing substantial remaining sequence difficulty.
Where the evidence stops. These heatmaps diagnose the evaluated rollouts; they are not a module ablation or evidence that latent prediction caused the gains. Stage-reach observations are correlated within a trial, and each percentage comes from only twenty trials.
7. Analysis & limitations
7.1 What the evidence leaves open
The central future-to-action explanation conflicts with the explicit action-to-future attention direction and action-only deployment. Equation (19)'s arguments and the noise-stream self-attention description also leave the exact implementation of visibility insufficiently resolved. The report follows the explicit Section 4.4 rules without treating them as code-verified behavior. E02E05E06
No mechanism ablation removes future prediction or compares endpoint and dense-video targets within a matched architecture. No future-latent accuracy or physical-state validation is reported. Consequently, the policy comparison does not establish that future supervision causes the gains or that its representation captures sufficient physical state. E14
Four tasks on one robot in fixed scenes with predefined resets provide limited evidence about distribution shifts or cross-embodiment transfer. Common fine-tuning data and evaluation rules do not control differences in pretrained models. Training-cost savings are asserted without a quantitative training-compute comparison. E02E07E14
The authors describe retaining popcorn while reopening an intervened microwave door, and using the held package to widen a narrow opening. They present these as preliminary adaptive behaviors rather than a quantitative emergence result; intervention counts, baseline comparisons, and success frequencies are absent. E13
7.2 Questions for discussion
- Does future-latent supervision improve action learning when the future-to-action attention path is absent, and how much of the gain survives λ = 0? [E05, E06, E14]
- How are the noise-stream preprocessing and Equation (19) reconciled with Section 4.4's visibility rules? [E05, E06]
- Would endpoint prediction retain its advantage under randomized object poses, contact conditions, and intermediate interventions evaluated quantitatively? [E07, E13]
8. Reproducibility audit
8.1 Requirements and known gaps
The supplied artifact is arXiv:2608.22067v4. Its title and all eight catalog authors match, but the title page additionally lists Cong Fang. This is an older or incomplete catalog author list; the recorded metadata is preserved. E01
Implementation requires the exact Qwen3 and DINO-v3 variants, trainable/frozen module choices, future offset Δt, token layouts, attention-mask scope, proprioception integration, and λ. These details are not sufficiently specified to reconstruct the reported model faithfully. E03E04E05E06
Evaluation replication also needs demonstration counts and splits, the proposed policy's optimization settings, reset poses, sampling configuration, and chunk-execution schedule. Preserve the ordered terminal-state rules and distinguish core latency from physical execution time. No code was inspected and no experiments were reproduced in this review. E06E07E08E10E11
A minimal mechanism test would compare the reported objective against λ = 0 using matched initialization, data, action architecture, and deployment settings. Measure physical success and stage retention across repeated training runs, and verify that future tokens cannot influence action outputs under the intended mask. E05E06E14
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: Does every implemented attention block respect the stated action mask?
Proposed check: hold instruction, current observations, proprioception, noisy actions, and random seeds fixed. During a training-mode forward pass, replace or permute only future-observation tokens and measure the change in action velocity outputs and their Jacobian with respect to future inputs. Repeat after removing the future groups as in deployment, auditing intermediate as well as final attention blocks. Under Figure 3’s claimed dependency, action outputs should be invariant up to numerical tolerance. A measurable dependence would expose a mismatch requiring resolution before attributing behavior to the reported mask. E04E05E06
Check 2: Does future-state supervision improve control under matched action visibility?
Proposed check: train matched variants with the stated action-first mask using the full latent objective, λ = 0 action-only training, and temporally shuffled future-latent targets. Hold encoders, initialization, demonstration split, optimization budget, action horizon, and inference steps constant. Evaluate the same reset configurations and terminal predicates, reporting full success, normalized progress, paired stage failures, and core/end-to-end latency separately. If true future targets outperform both controls, they support predictive supervision beyond capacity or extra training. If not, the baseline advantage alone would not establish the proposed world-modeling mechanism. E05E06E07E08E09E10E14
8.3 Reading coverage
Visual audit: Original PDF pages 7, 9, 12, and 13 and all five final crops were visually inspected. This covers Figure 2 stream architecture, Figure 3 mask orientation and train/inference difference, Table 2 numerical results, Figure 5 timing/performance axes, and Figure 6 complete stage-reach maps. The source is arXiv 2608.22067v4; its PDF SHA-256 was recomputed and matches the base report. The action/future mask was checked against Section 4.4 while preserving the unresolved Eq. 19 argument-list ambiguity. Other figures, external videos, implementation masks, and experimental reproduction remain outside this visual pass.
PDF pages inspected for this edition: 7, 9, 12, 13. Appendix coverage: not present.
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 page and abstract, PDF p. 1
- Contents, PDF p. 2
- Sections 1.1–1.2: limitations of video-based WAMs and proposed formulation, PDF pp. 3–4
- Sections 2.1–2.2: related work, PDF pp. 4–6
- Sections 3.1–3.3: manipulation, tokenization, and timestep embedding, PDF p. 6
- Sections 4.1–4.4: architecture, flow matching, training, and inference, PDF pp. 7–10
- Sections 5.1–5.5: setup, results, stage analysis, failures, and intervention observations, PDF pp. 10–14
- Section 6: conclusion, PDF p. 14
- References, PDF pp. 15–19
- Supplied textual captions for Figures 1–6
Outside the original text pass
- Figure images were not visually inspected; only supplied text and captions were read. Numerical coordinates from extracted plots were not reconstructed.
- No appendix appears in the supplied paper. Separate supplemental material availability remains unverified.
- Referenced publications and the linked project page were not independently inspected.
- 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
E01PDF p. 1, title page and arXiv version line
The title page identifies arXiv:2608.22067v4, dated 31 August 2026, and lists all eight catalog authors plus Cong Fang.
Go to primary source ↓E02PDF pp. 3–4, Sections 1.1–1.2
The authors motivate endpoint-state prediction by the dimensional and computational burden of video trajectories and describe actions as inferred from future states.
Go to primary source ↓E03PDF p. 6, Sections 3.1–3.2, Equations (1)–(2)
The formulation specifies H = 60, seven-dimensional single-arm and fourteen-dimensional dual-arm actions, three camera views, proprioception, Qwen3 language encoding, and DINO-v3 vision encoding.
Go to primary source ↓E04PDF pp. 7–8, Sections 4.1.1–4.1.3, Equations (6)–(13)
Language and vision are projected to width 1024, concatenated for self-attention, and transformed into normalized QKV features using condition-stream timestep 1.
Go to primary source ↓E05PDF pp. 8–9, Sections 4.2–4.3, Equations (14)–(21)
Action and future-image latents receive interpolated Gaussian noise. Stream QKV tensors are concatenated for masked attention. Action and observation velocity losses are combined with λ; Equation (19) includes the noisy future latent among action-predictor arguments.
Go to primary source ↓E06PDF pp. 9–10, Section 4.4; PDF p. 9, Figure 3 textual caption
Condition tokens cannot access targets; action tokens access context and actions but not future observations; future tokens access every group. Future-observation groups are removed at inference.
Go to primary source ↓E07PDF pp. 10–11, Section 5.1, robot/tasks, baselines, and evaluation protocol
Eight methods use one Astribot S1, fixed scenes, common task data and reset conditions. Baselines train for three equivalent epochs. Twenty trials per method-task pair yield 640 trials, with a 180-second limit and unsafe terminations counted as failures.
Go to primary source ↓E08PDF p. 10, Table 1; PDF p. 11, Section 5.1, Metrics
Door, Pepsi, Add Ice, and microwave tasks have four, five, six, and five ordered stages with explicit physical terminal conditions. Progress averages consecutive-stage fractions equally across tasks; Add Ice remaps eight raw stages into six semantic stages.
Go to primary source ↓E09PDF p. 12, Table 2, DELE-w0.5 and baseline rows; PDF p. 11, Section 5.2
DELE-w0.5 reports 81.3% macro progress and 62.5% overall success. Its task successes are 80%, 65%, 45%, and 60%. XR0 leads baseline overall success at 30%; GWP0.5 leads baseline macro progress at 61.3%. The prose gives improvements of 32.5 and 20.1 percentage points.
Go to primary source ↓E10PDF pp. 11–12, Section 5.2, latency discussion; PDF p. 12, Figure 5 textual caption
The reported median core latency is 87.5 ms on RTX 4090. The prose provides qualitative baseline comparisons and excludes transfer, decoding, action projection, IK, trajectory processing, and robot execution.
Go to primary source ↓E11PDF p. 12, Section 5.2, final paragraph
For 16, 13, 9, and 12 successful trials, mean task times are respectively 16.66, 45.92, 46.27, and 39.50 seconds. Times are conditional on successful completion.
Go to primary source ↓E12PDF pp. 13–14, Sections 5.3–5.4
The prose reports task completion counts and identifies coordination, transfer, ice acquisition and pouring, insertion, closure, and final-state maintenance as major bottlenecks.
Go to primary source ↓E13PDF p. 14, Section 5.5
Additional intervention rollouts qualitatively show reopening a closed microwave while retaining popcorn and pushing a narrow door opening with the package. The authors explicitly limit these observations to preliminary evidence.
Go to primary source ↓E14PDF pp. 10–14, Sections 5.1–5.5
The experimental section contains policy comparisons, latency and completion timing, stage/failure analyses, and intervention observations. It reports no future-loss ablation, matched endpoint/video-target comparison, future-latent prediction metric, or quantitative training-compute comparison.
Go to primary source ↓iv01PDF p. 7, Figure 2; original page and final crop visually inspected at 200 DPI
Figure 2 was visually inspected from language/current-state inputs through the condition stream and from action/next-state inputs through the noise stream. The fixed condition timestep, sampled noise timestep, AdaLN/QKV coupling, masked attention, and two output heads remain visible in the crop.
Go to primary source ↓iv02PDF p. 9, Figure 3; original page and final crop visually inspected at 200 DPI
Figure 3 was visually inspected directly: training A_t query versus O_(t+Δt) key/value is white; the reverse cell is colored; language/current-observation queries cannot read either prediction group. Inference omits O_(t+Δt). Section 4.4 confirms this direction, while Eq. 19 retains the future latent in the action-function argument list.
Go to primary source ↓iv03PDF p. 12, Table 2; original page and final crop visually inspected at 200 DPI
Table 2 was visually inspected with all eight method rows and paired progress/success columns. DELE’s per-task success is 80/65/45/60, macro progress 81.3, overall success 62.5; XR0 overall success is 30.0 and GWP0.5 macro progress is 61.3. The setup specifies twenty trials per method/task.
Go to primary source ↓iv04PDF p. 12, Figure 5; original page and final crop visually inspected at 200 DPI
Figure 5 panels (a)–(f), axis labels, method labels, and highlighted DELE points were visually inspected. Panel (b) is core-model latency; panels (c)–(f) explicitly condition completion time on success. Source prose reports 87.5 ms median core latency and lists the excluded runtime components.
Go to primary source ↓iv05PDF p. 13, Figure 6; original page and final crop visually inspected at 200 DPI
Figure 6 was visually inspected with the shared scale, all method labels, task blocks, and Full columns. DELE rows are Door 100/100/100/80; Pepsi 100/90/90/65/65; Add Ice 95/70/70/55/45/45; Microwave 100/100/85/80/60. These are ordered-prefix reach probabilities, not conditional transition probabilities.
Go to primary source ↓8.5 Primary sources
DELE-w0.5: Inferring Action from Future Latent State for Robotic Manipulation ↗
PDF · 10,239 extracted words
Source fingerprint
57eb52707dede1fa6bbbe32036a4bedc12d7871ab35c1465b7be005f21453550