Dyna-2: A 1-million-hour scaling law for world-action models

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

Authors: Dyna Robotics

Source: ref-ed0e9bb8027f431c1f20 ↗ · Project page ↗ · Catalog record

Reading: 276 / 558 · 6 original figures & tables · ~22 min ·

1. Paper overview

In one sentence: Human-video supervision supports cross-embodiment transfer while the scaling policy remains reactive; the reported gains depend on data scale, model variant and evaluation metric. e-problem e-data e-objective e-offline e-onrobot e-ablation e-vla e-language e-sampler

At a glance	What to know
Research problem	Source description Dyna asks whether increasing human manipulation experience improves held-out human prediction, transfers to robot data without robot pre-training, and helps physical execution after robot fine-tuning. The practical data problem is that reliable hand-pose labels lag behind available human video; the study tests whether future-video supervision can exploit that unlabeled experience. e-problem e-video-scale
Core mechanism	Source description A nested human-data ladder tests scaling from 1,000 to 1,000,000 hours, separating held-out human prediction, zero-shot offline robot prediction and post-trained physical execution. e-data e-human e-offline e-onrobot
A key reported result	Post-trained physical manipulation across 14 tasks: Reported mean normalized scores: 20% → 28% → 45% → 53%. Mean normalized task score ↑. Three embodiments; ten blind trials per task, twelve for language following; each native metric normalized to its task ceiling. 1k → 10k → 100k → 1M human hours; matched robot post-training. The aggregate averages fractions of each task's ceiling and is not a binary success rate. Individual tasks need not improve: Rope Tie falls from 90% at 100k to 40% at 1M, while Lockbox Key Turning rises from 0% to 90%. The source does not define the aggregate whiskers' uncertainty estimator. e-onrobot
Reading caution	Reader analysis Four data budgets and a limited robot suite support empirical scaling trends, not an unlimited scaling law or universal embodiment transfer. Physical trials are few and some tasks regress. Checkpoint variation is not independent-seed uncertainty. e-data e-human e-offline e-onrobot

Core contributions

Source description
A nested human-data ladder tests scaling from 1,000 to 1,000,000 hours, separating held-out human prediction, zero-shot offline robot prediction and post-trained physical execution. e-data e-human e-offline e-onrobot
Author claim
The authors identify additional video-only training as a driver of cross-embodiment transfer and introduce a separate adaptive-target distillation method for one-step manipulation-video generation. e-video-scale e-distillation

A green video transformer stack receives context frames, noisy future video and text cross-attention. A shallower blue action stack receives proprioception and noisy actions. The solid video-to-action arrow carries observed context; the dashed future-video arrow is labeled L_joint only. — **Figure 3.** Observed context supports reactive action prediction; the future-video arrow belongs to a separately labeled variant. Original source, #fig-3 ↗
Extracted original HTML/SVG content with original author fonts and a documented isolated layout wrapper; original chart data, geometry, text, captions, legends and footnotes are retained. The wrapper expands the carousel, reserves native SVG aspect-ratio height and displays native source MathML while hiding its duplicate KaTeX HTML representation. This is not a pixel-identical historical website capture or a publisher PDF.

2. Motivation

2.1 The problem and the proposed response

Source description

Dyna asks whether increasing human manipulation experience improves held-out human prediction, transfers to robot data without robot pre-training, and helps physical execution after robot fine-tuning. The practical data problem is that reliable hand-pose labels lag behind available human video; the study tests whether future-video supervision can exploit that unlabeled experience. e-problem e-video-scale

2.2 What this reading follows

Dyna-2 asks whether human manipulation experience can become useful robot pre-training. Human hand tracks provide pseudo-actions, while future-video prediction offers supervision even when reliable action labels are unavailable. The central experiment increases human-data budgets and evaluates both held-out human actions and robot actions, then separately tests robot policies after task post-training. This reading follows that separation through the original architecture, robot results and objective ablation. It also examines production comparisons, partial-credit instruction scores and one-step video sampling. The source reports encouraging mean trends, but those trends do not establish universal task improvement, significant superiority in every comparison, or future-video planning during reactive control. e-problem e-data e-objective e-offline e-onrobot e-ablation e-vla e-language e-sampler

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	Other mechanisms
Quadrant	Outside quadrants

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

The recorded WAM, Other mechanisms and Outside quadrants labels are compatible with L_co's auxiliary future-video training and reactive action inference. Architecture depends on the model boundary: the authors describe one integrated generative mixture of transformers, yet video and action have distinct DiT stacks, with a shallower action expert joining early video layers. Counting modality experts can support a Dual-system interpretation; counting the integrated generator can support One Model. Neither the authors' 'single model' phrase nor co-training alone settles the catalog's expert-boundary convention. No separate planner/controller deployment is demonstrated for the scaling variant, and the L_joint-only arrow must not be generalized to it. Preserve the recorded snapshot rather than force a reclassification. e-architecture e-objective

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
Past video frames, proprioception and language instruction Training: future video and human hand-pose-derived pseudo-actions; robot demonstrations for post-training	Future action chunks for the reactive policy Future-video latents under the video objective and separate generation configurations

4.2 Equations and their role

z_t=t z+(1-t)\varepsilon_z,\qquad a_t=t a+(1-t)\varepsilon_a,\qquad \varepsilon_z,\varepsilon_a\sim\mathcal{N}(0,I)

The source interpolates future-video latent z and future action chunk a with Gaussian noise. The interpolation parameter t defines corrupted inputs; the target velocities are z−ε_z and a−ε_a. Conditioning c consists of past frames, proprioception and instruction. e-objective

\mathcal{L}_{\mathrm{co}}(\theta)=\mathbb{E}\left\|u_\theta^{\mathrm{vid}}(z_t;t,c)-(z-\varepsilon_z)\right\|^2+\lambda\mathbb{E}\left\|u_\theta^{\mathrm{act}}(a_t;t,c)-(a-\varepsilon_a)\right\|^2

The network u_θ predicts modality-specific velocities; λ weights the action term. Its numerical value is unspecified. The absence of z_t from the action field is the explicit reason that co-training does not require future-video generation during control. e-objective

5. Method in detail

5.1 Separate representation learning from future-conditioned control

Reader analysis

The action supervision begins with quality-controlled human hand tracks: wrist poses supply end-effector trajectories, and thumb–index aperture supplies a continuous grasp signal. Video and action then enter distinct transformer stacks joined by attention. In the flow-matching equations, c denotes observed frames, proprioception and instruction; z and a denote future-video latents and an action chunk. The L_co loss fits two marginal velocity fields. Its action term receives a_t, t and c, with no z_t argument, so the source explicitly keeps scaling-model inference reactive. Reader interpretation: predicting the world during training can still help the representation used for action. Figure 3's dashed L_joint-only arrow depicts an additional conditional path, whose detailed objective and relationship to the joint ablation remain unspecified. e-data e-architecture e-objective e-ablation

5.2 Keep the three transfer tests on separate axes

Reader analysis

The data ladder contains nested 1k, 10k, 100k and 1M-hour subsets with fixed source proportions. A disjoint 100-hour human validation set first tests in-domain action prediction. A separate 39-task robot suite then tests offline transfer without robot pre-training or adaptation. These accuracy metrics count action dimensions within normalized thresholds; they do not count completed tasks. Physical execution comes later, after matched robot-only post-training on 14 tasks with at most ten hours of demonstrations per task. Its reported 20/28/45/53% curve averages task scores divided by their ceilings. Reader interpretation: improvements across these three settings are complementary, but not interchangeable. Figure 8's task regressions and the small blind-trial counts prevent interpreting the rising mean as uniformly reliable execution. e-data e-metrics e-human e-offline e-onrobot

5.3 Use the ablation to narrow the mechanism claim

Reader analysis

At matched action-data budgets, the joint recipe adds video prediction and the added-video recipe also adds unlabeled human exposure. Extra video loses to joint at 5k hours before outperforming it at larger budgets. The fixed-action sweeps strengthen the transfer interpretation: more video improves robot prediction while human prediction does not similarly improve. Reader interpretation: this supports a useful cross-embodiment training signal, but changed exposure and the zero-video action-only baseline complicate an objective-only causal claim. Keep Section 4 separate, because its production recipes differ. The early WAM/VLA comparison has intervals crossing equality; the instruction benchmark gives half credit for failed correct attempts; and one-step distillation measures video-sampling speed with reduced motion quality. None establishes a future-conditioned controller for the scaling variant. e-ablation e-video-scale e-variants e-vla e-language e-sampler e-objective

5.4 Training and inference

During training

Source description

The scaling model fits separate video and action velocity fields with a shared trunk. Nested data budgets preserve source proportions; a fixed, disjoint 100-hour human validation set is used. The source states identical training/evaluation configurations and defers compute/model-size scaling. e-objective e-data

Source description

Physical evaluation post-trains equivalently stepped checkpoints on the same robot-only task datasets, each at most ten hours, without human–robot alignment or co-training. Section 4 uses different production recipes; its early WAM/VLA comparison instead uses action-only supervision and predates million-hour pre-training. e-onrobot e-variants e-vla

During inference

Reader analysis

For the L_co scaling variant, action denoising takes observed context and a noisy action chunk, without noisy or predicted future video. The text explicitly says the policy neither generates nor attends to predicted futures during inference. Thus video prediction supplies training supervision here, not a demonstrated rollout planner or inverse-dynamics stage. Physical action execution is assessed after robot post-training, separately from offline transfer. e-architecture e-objective e-onrobot

5.5 Implementation flow

Extract action supervision from human activity
Quality-controlled 3D hand tracks supply wrist poses as end-effector trajectories and thumb–index aperture as a continuous grasp signal. Video without reliable action labels can still train future prediction. The authors report no embodiment-specific visual or kinematic alignment preprocessing. e-data e-video-scale
Connect modality-specific transformers
The source calls Dyna-2 one generative mixture of transformers, with distinct video and action DiT stacks connected by attention. The shallower action stack joins early video layers, receives proprioception and attends to observed-video context. Video masking is causal; action self-attention is bidirectional. Text conditions video through cross-attention, with no direct text-to-action path. Figure 3 separately marks its dashed future-video-to-action arrow as L_joint only; it is not the reactive L_co scaling-model path. e-architecture
Distill a separate video generator
A one-step student follows a smoothed target-distribution path from an initially reachable target toward data. Student updates are fast; target advancement is slower and gated by an online student readout. The source leaves the divergence choice open and does not fully specify the controller. e-distillation

6. Experiments & results

Dyna-2 tests whether scaling human manipulation video improves robot learning. Its scaling variant co-trains future-video and action prediction through shared representations but remains reactive during control. The source reports improvements in offline robot prediction and in mean normalized robot performance after task post-training. Separate production experiments study VLA comparisons, deployment, language following and fast video generation. Six original HTML visuals explain these findings while preserving their uncertainty and variant boundaries.

Source and visual limitations

Reader analysis

The supplied inventory supports six static original HTML/SVG visuals, including architecture, physical results, an ablation and a quantitative sampler table. The unavailable external celery photographs, generation comparison image, rollout videos and generated-video mosaics limit independent assessment of visual quality and continuous behavior. Figure 3 is a static source schematic, not an animation-derived still. Reported generation metrics and robustness descriptions cannot substitute for viewing those media. e-architecture e-qualitative e-distillation e-sampler

6.1 Read the original evidence

The top plot reports mean normalized task scores of 20, 28, 45 and 53 percent across 1k, 10k, 100k and 1M human hours. Fourteen lower panels show native task metrics, including Rope Tie's decline from 90 to 40 percent at the largest scale. — **Figure 8.** The reported mean improves across data budgets, while individual tasks can regress. Original source, #fig-8 ↗
Extracted original HTML/SVG content with original author fonts and a documented isolated layout wrapper; original chart data, geometry, text, captions, legends and footnotes are retained. The wrapper expands the carousel, reserves native SVG aspect-ratio height and displays native source MathML while hiding its duplicate KaTeX HTML representation. This is not a pixel-identical historical website capture or a publisher PDF.

Early Dyna-2 has green bars labeled 1.55 times normalized performance and 1.12 times normalized grade against purple Dyna-1 bars at 1.00. Both green paired-bootstrap confidence intervals extend below and above the 1 times equality line. — **Figure 13.** Positive pooled mean ratios coexist with confidence intervals that include equality. Original source, #fig-13 ↗
Extracted original HTML/SVG content with original author fonts and a documented isolated layout wrapper; original chart data, geometry, text, captions, legends and footnotes are retained. The wrapper expands the carousel, reserves native SVG aspect-ratio height and displays native source MathML while hiding its duplicate KaTeX HTML representation. This is not a pixel-identical historical website capture or a publisher PDF.

Grouped bars show language-following scores for three training recipes. The aggregate values are 0.35, 0.67 and 0.96. The caption awards one point for completion, half a point for attempting the correct primitive without completion, and zero for a wrong primitive. — **Figure 15.** Video co-training improves a partial-credit language score, not a directly interchangeable completion rate. Original source, #fig-15 ↗
Extracted original HTML/SVG content with original author fonts and a documented isolated layout wrapper; original chart data, geometry, text, captions, legends and footnotes are retained. The wrapper expands the carousel, reserves native SVG aspect-ratio height and displays native source MathML while hiding its duplicate KaTeX HTML representation. This is not a pixel-identical historical website capture or a publisher PDF.

The full sampler table lists NFE, latency, speedup, FVD, motion and flicker for real video, the default and truncated teacher, DMD2 variants and the proposed student. The student row reports one evaluation, 110 milliseconds, 93 times speedup, FVD 121 and motion 75 percent. — **Video sampler comparison table.** One-step video sampling trades some teacher quality for much lower sampler latency. Original source, #fig-sampler-table ↗
Extracted original HTML/SVG content with original author fonts and a documented isolated layout wrapper; original chart data, geometry, text, captions, legends and footnotes are retained. The wrapper expands the carousel, reserves native SVG aspect-ratio height and displays native source MathML while hiding its duplicate KaTeX HTML representation. This is not a pixel-identical historical website capture or a publisher PDF.

6.2 Results and evaluation conditions

Task & protocol	Reported result	Comparison & interpretation
Held-out human action prediction Fixed 100-hour human validation set; four nested pre-training budgets; ten late-window checkpoints summarized by mean and standard deviation.	Figure 5: MSE 0.062 → 0.057 → 0.056 → 0.054; L1 0.140 → 0.131 → 0.129 → 0.127; accuracy@0.1 0.017 → 0.021 → 0.024 → 0.026; accuracy@0.5 0.40 → 0.44 → 0.45 → 0.47. Action MSE ↓; L1 ↓; dimension-wise accuracy@0.1 and @0.5 ↑	Budgets: 1k, 10k, 100k, 1M hours, respectively. These rounded reported means improve monotonically. Accuracy concerns action dimensions, not completed episodes. Checkpoint-variation whiskers are visible, but numerical standard deviations are not tabulated; variation across ten late checkpoints is not independent-seed uncertainty. e-data e-metrics e-human
Zero-shot offline robot action prediction 39 held-out tasks: 12 internal and 27 xdof ABC, on two stationary bimanual YAM platforms; no robot pre-training or adaptation.	Figure 2 labels: MSE 0.180 → 0.174 → 0.124 → 0.117; accuracy@0.5 0.067 → 0.074 → 0.136 → 0.159. Action MSE ↓; accuracy@0.5 ↑	Same four human-data budgets. Offline predictive transfer across the evaluated embodiment gap; these are not zero-shot robot execution rates. e-overview-values e-metrics e-offline
Post-trained physical manipulation across 14 tasks Three embodiments; ten blind trials per task, twelve for language following; each native metric normalized to its task ceiling.	Reported mean normalized scores: 20% → 28% → 45% → 53%. Mean normalized task score ↑	1k → 10k → 100k → 1M human hours; matched robot post-training. The aggregate averages fractions of each task's ceiling and is not a binary success rate. Individual tasks need not improve: Rope Tie falls from 90% at 100k to 40% at 1M, while Lockbox Key Turning rises from 0% to 90%. The source does not define the aggregate whiskers' uncertainty estimator. e-onrobot
Objective and video-only-data ablations Matched-step offline evaluation on the 39-task robot suite.	The authors report joint beating action-only on 39/39 tasks at each of 5k, 50k and 100k action hours. Added video is worse than joint at 5k on both plotted metrics, but better at 50k and 100k. Reported taskwise advantage; robot action MSE trend	Video-only sweeps: 0/1k/10k/50k hours at fixed 50k action hours; 0/250k/750k at fixed 250k action hours. The negative 5k condition prevents an unconditional claim that extra video helps. The added-video recipe improves with scale in this comparison; action-only and joint do not improve monotonically. More video also changes data exposure, so this is not a pure objective-only intervention. e-ablation e-video-scale
Video-only scaling with fixed action supervision Offline held-out robot evaluation; action-labeled data fixed separately at 50k or 250k human hours.	Figure 11 labels robot MSE 0.34 → 0.12 across 0 → 50k video-only hours at 50k action hours, and 0.10 → 0.084 across 0 → 750k video-only hours at 250k action hours. Figure 12 labels the final human and robot errors 104% and 34% of their respective zero-video baselines. Robot action MSE ↓; domain-relative error	Video-only sweeps are 0/1k/10k/50k and 0/250k/750k hours, respectively. The source attributes the benefit to cross-embodiment generalization; human error slightly worsens. The zero-video arm is described as action-only, so its comparison also changes the training objective. The proposed gradient-dilution explanation is an author hypothesis, not a measured mechanism. e-video-scale e-offline e-metrics
Early Dyna-2 versus Dyna-1 on-robot comparison Seven tasks × three pre-training checkpoints; matched datasets and training hyperparameters.	1.55× success rate; 1.12× grade. Pooled success-rate ratio; quality-grade ratio	Dyna-1 VLA normalized to 1.00; early action-only Dyna-2 WAM. These are ratios of pooled means over 21 task×checkpoint cells, not means of per-cell ratios. Both plotted paired-bootstrap 95% confidence intervals cross 1, so the intervals do not establish superiority at that level. Exact endpoints are not labeled. This early action-only comparison does not isolate the final video co-training objective. e-vla
Production pass criteria at unseen customer sites Same task post-training datasets and steps; neither model saw deployment-site data; operator assessment against customer criteria.	Dyna-2 87%; Dyna-1 46%. Production pass rate ↑	41 percentage points; both approximately 100% in house. Site generalization after task post-training, not human-only zero-shot control. Trial counts and detailed acceptance rubrics are unspecified. e-deployment
Counterfactual instruction following Four instruction-following tasks: push/pull Jenga, object kitting, piece stacking and napkin manipulation; Figure 15 labels n=36 in total, with task counts 8/10/10/8. Each model is fine-tuned on a few hours of task data.	0.35 → 0.67 → 0.96. Language-following score, 0–1 ↑	Early action-only → early video co-training → full-corpus video co-training. Half credit is awarded when the commanded primitive is attempted but fails. Despite prose calling these success rates, 0.96 is a partial-credit score, not 96% completed tasks. e-language
One-step manipulation-video generation Held-out episodes; three-second, three-view output; sampler latency on one H100.	Student: 110 ms, FVD 121, motion 75%; default teacher: 10,203 ms, FVD 80, motion 94%. Sampler latency ↓; FVD ↓; motion relative to real	One versus 100 network evaluations; table reports 93× speedup. This measures a video-sampling speed–quality tradeoff, not action-control latency or planning success. Student flicker is 1.94 versus real 2.37 and teacher 2.69; low flicker alone does not establish correct motion. The table's 'Teacher, steps cut to 1' row reports NFE=2, an unresolved distinction between the row label and evaluation count. End-to-end decoding time is not established. e-distillation e-sampler

6.3 Ablations and diagnostic examples

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

Two plots compare action-only, joint and added-video recipes at 5k, 50k and 100k action-labeled hours. Added video has worse MSE and accuracy than joint at 5k, then better values at the larger budgets. — **Figure 10.** Extra video helps at larger budgets, but the 5k condition is a negative result. Original source, #fig-10 ↗
Extracted original HTML/SVG content with original author fonts and a documented isolated layout wrapper; original chart data, geometry, text, captions, legends and footnotes are retained. The wrapper expands the carousel, reserves native SVG aspect-ratio height and displays native source MathML while hiding its duplicate KaTeX HTML representation. This is not a pixel-identical historical website capture or a publisher PDF.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Four data budgets and a limited robot suite support empirical scaling trends, not an unlimited scaling law or universal embodiment transfer. Physical trials are few and some tasks regress. Checkpoint variation is not independent-seed uncertainty. e-data e-human e-offline e-onrobot

Reader analysis

Figure 3 labels the dashed predicted-future-to-action arrow L_joint only, whereas Section 2 formalizes L_co and explicitly excludes predicted futures from the scaling policy's inference. Section 3 also uses the word joint for an ablation recipe without supplying a corresponding L_joint equation or fully specifying its relation to that arrow. Preserve this variant ambiguity; do not assign future-conditioned control to the scaling results. e-architecture e-objective e-ablation

Reader analysis

Production recipes differ from the scaling models. Disturbance robustness is described through selected demonstrations, without quantitative disturbance trials; the source expressly distinguishes sensor-loss robustness from prediction of unseen state. e-variants e-qualitative

Reader analysis

The production comparison's positive mean ratios coexist with confidence intervals crossing equality. Language-following prose calls partial-credit scores success rates, and the sampler's one-step teacher label coexists with NFE=2. These are distinct evidential or reporting issues; no unseen implementation is invoked to resolve them. e-vla e-language e-sampler

7.2 Questions for discussion

Which held-out embodiment would most strongly test whether the observed transfer extends beyond the YAM-centered evaluation?
Does the one-step student's lower motion score compromise a downstream planner despite its much lower sampling latency?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproduction needs the corpus split and sampling rules, hand-pose quality thresholds, action normalization, architecture dimensions, loss weight, optimizer schedule, action horizon and execution cadence. These are not fully specified. Distillation additionally needs target-path/readout/controller definitions and metric implementations; the H100 sampling measurement is not a training-compute budget. e-data e-metrics e-architecture e-objective e-distillation e-sampler

Reader analysis

Proposed check: hold action-labeled examples, architecture, optimizer updates and compute budget fixed; compare action-only, video prediction on those same examples, and extra video-only exposure, with matched-exposure controls and multiple seeds. Evaluate both human and robot splits. A transfer gain specific to video prediction would strengthen the claimed mechanism. e-ablation e-video-scale

Reader analysis

Proposed check: repeat the four-budget robot post-training comparison with identical demonstrations and randomized starts, more blind trials and taskwise confidence intervals. Publish binary success separately from normalized scores. Test whether the aggregate gain survives uncertainty while Rope Tie still regresses; this distinguishes average transfer from universal improvement. e-onrobot

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: Separate video prediction from additional exposure

Reader-proposed experiment: hold action-labeled examples, architecture, optimizer updates and compute fixed. Compare action-only, video prediction on those same examples, and extra video-only exposure with matched-exposure controls and multiple seeds. Retain the 5k condition instead of testing only scales where added video wins. Evaluate both held-out human and robot action metrics. A robot-specific improvement that survives these controls would strengthen the transfer mechanism; a disappearing gain would point to exposure or optimization effects. e-ablation e-video-scale e-metrics

Check 2: Repeat physical transfer with explicit aggregation and uncertainty

Reader-proposed experiment: repeat the four-budget post-training study with identical demonstrations, randomized starts, more blind trials and taskwise confidence intervals. Publish raw counts, binary completion and ceiling-normalized scores separately, and specify the aggregate uncertainty estimator. Check whether the mean improvement remains while Rope Tie or other tasks regress. Also define whether the catalog counts modality experts or the integrated generator before assigning One Model or Dual-system; this classification check is distinct from measuring transfer. e-data e-onrobot e-architecture e-objective

8.3 Reading coverage

Visual audit: All ten primary-text chunks were read, including repeated responsive content. The original identity image and supporting excerpts support-01 through support-49 were delivered by the supplied helper and actually viewed. All six final copied PNGs were separately viewed. Figure 3's arrow labels were checked against the L_co equation and reactive-inference paragraph; Figure 8's full expanded task panels, Figure 10's negative 5k condition, Figure 13's confidence intervals crossing 1, Figure 15's partial-credit caption and the complete sampler table were inspected. Original captions, legends and table notes are retained. Images use the pinned isolated HTML layout wrapper, not historical page styling. External photos and videos remain uninspected; no animation playback, independent rerender or reviewer approval is claimed by this writer. During the identity repair, all ten chunks and all fifty supporting images were read or viewed again, and the six unchanged final assets were recopied from the helper delivery and viewed. Metadata preserves the pinned primary title string, including its line break.

Original HTML supporting sections inspected: identity, support-01, support-02, support-03, support-04, support-05, support-06, support-07, support-08, support-09, support-10, support-11, support-12, support-13, support-14, support-15, support-16, support-17, support-18, support-19, support-20, support-21, support-22, support-23, support-24, support-25, support-26, support-27, support-28, support-29, support-30, support-31, support-32, support-33, support-34, support-35, support-36, support-37, support-38, support-39, support-40, support-41, support-42, support-43, support-44, support-45, support-46, support-47, support-48, support-49. 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

Research title, author/date fields and introductory figure captions
§ 1 Introduction
§ 2 Model architecture and training objectives
§ 3 The manipulation scaling laws: From 1,000 to 1,000,000 hours — data, metrics, human and robot scaling, post-training, objective ablations and video-only scaling
§ 4 Additional Model Capabilities and Extended Analysis — WAM/VLA comparison, qualitative cases, deployment, language following and one-step distillation
§ 5 Related Work
§ 6 Conclusion
§ References and Citing this research

Outside the original text pass

External figure assets are not downloaded; HTML alt text and mathematical text are retained where available.
The source HTML repeats responsive-page passages; repeated passages are not independent evidence.
The acquisition manifest records that external figure assets were not downloaded and that responsive passages repeat. Both responsive copies were read in all ten supplied text chunks; duplicates are not independent evidence.
The pinned upgrade supplies static original HTML/SVG figures and an HTML table under wam-original-html-evidence-v1. These were visually inspected with the supporting excerpts. No primary PDF or separate appendix was supplied. The isolated wrapper reflows the source, expands the carousel and retains captions, legends, footnotes and MathML; these images are not historical website screenshots.
Unavailable external images remain uninspected: the celery comparison photographs and Figure 16's generation comparison. Unavailable videos remain uninspected: Figures 1, 4, 6, 9, 17 and 18 and the lighting, sensor-loss and persistence demonstrations. Reading their captions does not establish their visual content, continuous behavior or execution success.
The supplied descriptor has no original-raster or animation-derived-still media items, component-time disclosure, model-rate annotations or state-field interface. No animation playback, rate measurement or media derivation is claimed.
The original identity section explicitly states Author: Dyna Robotics and Date: August 2026. This is a verified corporate byline, not authorship inferred from branding or copyright. No individual author, affiliation, publication day or numbered revision is stated. The title matches the catalog after capitalization and line-break normalization.
The extracted animated corpus counters do not establish usable clip, instruction or object counts. Those statistics remain unverified.
Referenced publications, code, weights and datasets were not inspected. No source scripts or paper code were executed, and no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityOriginal identity section: title, Author and Date fields (identity)Inspect

The original header names Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models, explicitly credits Author: Dyna Robotics, and dates the resource August 2026.

Go to primary source ↓

e-problem§ 3 opening research questions (support-11), Pre-training data (support-13), and Video is the new scaling axis (support-26)Inspect

The source asks about human scaling, zero-shot offline robot transfer, fine-tuned physical performance and the conditions enabling transfer. It describes a human-video corpus and a bottleneck in obtaining reliable action annotations.

Go to primary source ↓

e-architecture§ 2 architecture paragraphs (support-04, support-06) and Figure 3 (support-05; #fig-3)Inspect

The text describes one generative mixture of transformers with modality-specific DiT layers, observed-video attention from actions, causal video masking, bidirectional action attention, video tokens cross-attending to text and shallower early-joining action layers. Figure 3 shows a solid observed-context arrow to actions and a separate dashed predicted-future arrow labeled L_joint only.

Go to primary source ↓

e-objective§ 2 conditioning definitions and straight-path corruption (support-06, support-07), velocity-field setup (support-08), L_co equation (support-09), and reactive-inference paragraph (support-10)Inspect

The co-training fields share a trunk but fit separate marginals. The action velocity field takes noisy actions, time and context, without future-video input; the text explicitly excludes generating or attending to predicted future video during scaling-model inference.

Go to primary source ↓

e-data§ 3 Pre-training data: nested, exact-hours subsets of human experience (support-13, support-14)Inspect

Wrist poses and thumb–index aperture derive pseudo-actions from quality-controlled human hand tracks. Nested 1k/10k/100k/1M-hour subsets preserve source proportions, with a disjoint 100-hour human validation set. Embodiment-specific alignment is excluded; compute and model-size scaling are deferred.

Go to primary source ↓

e-metrics§ 3 Evaluation metrics, action-dimension/horizon averaging and normalized thresholds (support-15)Inspect

MSE and L1 average over action dimensions and chunk horizon. Accuracy is the fraction of normalized action dimensions within thresholds 0.5 or 0.1, not task success.

Go to primary source ↓

e-human§ 3 human scaling protocol (support-16) and Figure 5's four panels (support-17; #fig-5)Inspect

Ten late-window checkpoints supply reported means and standard deviations. Figure 5 labels MSE 0.062/0.057/0.056/0.054, L1 0.140/0.131/0.129/0.127, accuracy@0.1 0.017/0.021/0.024/0.026 and accuracy@0.5 0.40/0.44/0.45/0.47 across the four budgets; whiskers are visible.

Go to primary source ↓

e-overview-valuesFigure 2, Zero-shot Robot Data panels (support-02; #fig-2)Inspect

The robot panels list MSE 0.180/0.174/0.124/0.117 and accuracy@0.5 0.067/0.074/0.136/0.159 at 1k/10k/100k/1M hours.

Go to primary source ↓

e-offline§ 3 human-to-robot offline evaluation (support-18) and Figure 7 (support-19; #fig-7)Inspect

The held-out robot suite contains 12 internal plus 27 xdof ABC tasks on two stationary bimanual YAM platforms. These checkpoints receive no robot training trajectories or adaptation.

Go to primary source ↓

e-onrobot§ 3 on-robot protocol (support-20), Figure 8 all panels (support-21; #fig-8), and result discussion (support-22)Inspect

Four pre-training budgets undergo matched robot-only post-training for 14 tasks, each with at most ten hours of robot data. Ten blind trials per task, twelve for language, span three embodiments. Ceiling-normalized aggregate scores are 20/28/45/53%; Rope Tie is 0/40/90/40%. At 100k/1M, Lockbox Key Turning is 0/90%. The task metrics include object counts divided by ceilings of six or ten, alongside task success fractions. Aggregate whiskers are drawn without an estimator definition in this section.

Go to primary source ↓

e-ablation§ 3 three-recipe definitions (support-23), matched-step comparison and negative 5k condition (support-24), Figure 10 (support-25; #fig-10)Inspect

At 5k/50k/100k action-labeled hours, action-only, joint and additional video co-training recipes share architecture and matched evaluation steps. The authors report joint beating action-only on all 39 tasks; only added video co-training improves consistently with scale. The text explicitly says extra video is not advantageous at 5k; the original Figure 10 shows higher MSE and lower accuracy for added video than joint at that rung.

Go to primary source ↓

e-video-scale§ 3 Video is the new scaling axis (support-26); Figure 11 (support-27; #fig-11) and Figure 12 (support-28; #fig-12)Inspect

Video-only data scales through 0/1k/10k/50k hours with 50k action hours, and 0/250k/750k with 250k action hours. Robot prediction improves, whereas held-out human prediction does not; extracting reliable action annotations is described as a bottleneck. Figure 11 labels robot MSE endpoints 0.34/0.12 and 0.10/0.084; Figure 12 labels final domain-relative error human 104% and robot 34%. The text describes the zero-video condition as action-only and offers gradient dilution only as a hypothesis.

Go to primary source ↓

e-variants§ 4 Additional Model Capabilities and Extended Analysis, opening production-variant paragraph (support-29)Inspect

The authors distinguish production-model training recipes from those used specifically for the scaling-law experiments.

Go to primary source ↓

e-vla§ 4 WAM vs. VLA protocol (support-30) and Figure 13 bars, whiskers and complete caption (support-31; #fig-13)Inspect

Early Dyna-2 uses action-only supervision and predates million-hour training. Seven tasks and three pre-training checkpoints yield pooled WAM/VLA ratios of 1.55 for success and 1.12 for grade. Caption defines a ratio of means and paired-bootstrap 95% intervals. The original whiskers cross the 1× equality line for both metrics; numerical interval endpoints are not printed. The source states matched pre/post-training data and hyperparameters, with a VLA-tuned recipe.

Go to primary source ↓

e-qualitative§ 4 Qualitative case study: celery quality (support-32), altered lighting (support-33), sensor-loss boundary (support-34), and persistence (support-35)Inspect

The authors describe improved chopping quality, tolerance of lighting changes, partial sensor loss and repeated disturbance. They explicitly avoid interpreting sensor-loss robustness as prediction over unobserved scene state.

Go to primary source ↓

e-deployment§ 4 Zero-shot real-world deployment protocol (support-36) and Figure 14 (support-37; #fig-14)Inspect

Matched task post-training is followed by evaluation at unseen customer sites. Operator-reported production pass criteria give 87% for Dyna-2 versus 46% for Dyna-1; in-house rates are described as close to 100%.

Go to primary source ↓

e-language§ 4 Instruction following rationale and fine-tuning protocol (support-38, support-39); Figure 15 and scoring caption (support-40; #fig-15); success-rate wording (support-41)Inspect

Figure 15 displays push/pull Jenga, object kitting, piece stacking and napkin manipulation, with n=8/10/10/8 and total n=36. Aggregate scores are 0.35/0.67/0.96 for early action-only, early video co-training and full-corpus video co-training. Correct but unsuccessful attempts receive 0.5, completed commands 1 and wrong primitives 0. Surrounding prose calls these success rates; the caption defines partial credit. Task fine-tuning uses a few hours.

Go to primary source ↓

e-distillation§ 4 One-Step Video Generation: smoothed target path (support-42), coupled student/target equations (support-43), and controller gate (support-44)Inspect

The student and target distributions are smoothed before comparison. A student-sample readout gates slow advancement of the target path while student parameters update faster. The divergence may be mode-seeking, adversarial or mixed; implementation choices remain open.

Go to primary source ↓

e-sampler§ 4 One-Step Video Generation, H100/three-second/three-view paragraph (support-45) and complete sampler table (support-46; #fig-sampler-table)Inspect

For three-second, three-view videos on one H100, the table reports default teacher NFE 100, 10,203 ms, FVD 80 and motion 94%; student NFE 1, 110 ms, speedup 93×, FVD 121 and motion 75%. This measures video sampling. Flicker is 1.94 for the student, 2.69 for the teacher and 2.37 for real video. 'Teacher, steps cut to 1' has NFE=2, sampler 210 ms and FVD 1039; no explanation of that count is provided.

Go to primary source ↓

8.5 Primary sources

Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models ↗

HTML · 16,994 extracted words

Source fingerprint

631a144dea3d7940a559194f32fbe09e2619a910da490441f187a6a076ce6034