World Action Models are Zero-shot Policies
1. Paper overview
In one sentence: Joint video-action denoising makes a pretrained video model usable as a generalizing robot policy, while closed-loop feedback and Flash training trade computation against execution quality. e02e03e06e10e12e14e16e21
| At a glance | What to know |
|---|---|
| Research problem | Source description The authors distinguish semantic recognition from executing unfamiliar motions. Their question is whether video pretraining supplies temporal priors for learning from heterogeneous robot demonstrations and acting in new environments, without repetitive demonstrations for every target task. e02e03 |
| Core mechanism | |
| A key reported result | AgiBot unseen-task generalization: 39.5% Average task progress (%). 10 held-out tasks, unseen environments and objects; eight rollouts per task across four robots (80 per checkpoint). Pretrained π0.5: 16.3%; pretrained GR00T N1.6: 5%. A 23.2-percentage-point gain over the strongest listed VLA. This is partial completion, not full success. e10e12 |
| Reading caution |
Core contributions
Figure 4. One transformer predicts both modalities; real observations refresh its history. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left half as a training graph. Video enters the VAE, actions enter an action encoder, and both noisy representations enter the same causal DiT. The joint flow-matching objective supervises both outputs. On the right, follow the action decoder into asynchronous real-world execution, then trace the green return path to observations and the KV cache. That path is essential: the next chunk is conditioned on what actually happened. Figure 14 specifies the attention pattern behind these blocks: the current video/action chunk can use previous clean conditioning frames, rather than previous action predictions. Equation (1) describes this coupled system as an implicit video-prediction/IDM factorization. e03e04e05e06e20
What it supports. The architecture supports the catalog’s One Model judgment through shared computation, not merely shared training data. Its inverse dynamics are implicit in joint denoising. Generated video helps form actions at inference, but the predicted frames are discarded from future history once real feedback becomes available.
Where the evidence stops. Refreshing history removes dependence on previously imagined frames; it does not ensure a correct current plan or successful execution. Appendix H illustrates actions following an incorrect visual plan. The paper also leaves memory-dependent task performance untested.
2. Motivation
2.1 The problem and the proposed response
The authors distinguish semantic recognition from executing unfamiliar motions. Their question is whether video pretraining supplies temporal priors for learning from heterogeneous robot demonstrations and acting in new environments, without repetitive demonstrations for every target task. e02e03
2.2 What this reading follows
DreamZero asks whether the motion knowledge in a video generator can become an executable robot policy. Its shared transformer predicts visual futures and action chunks together, while a controller carries out the actions and feeds real observations back into the model. The experiments make a useful distinction between unfamiliar environments, tasks absent from robot training, and tasks introduced through video-only transfer. This reading follows the information flow first, then separates these evaluation settings and examines the cost of fast inference. The strongest evidence concerns real robot task progress; it leaves important questions about full success, motion novelty, and reproducible training details. e02e03e06e10e12e14e16e21
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 | IDM |
| Quadrant | Q2 · One Model × IDM |
3.1 Evidence-based assessment
Supports the recorded classification
One Model is supported by the shared DiT and joint denoising. IDM is defensible as the paper’s implicit factorization, but the operational mechanism is joint future/action prediction, not an explicit video-then-IDM cascade. World prediction remains active at inference, beyond an auxiliary training loss. e03e04e06
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 Understand the implicit IDM inside joint prediction
Start from what the model must explain in a robot demonstration: how the observed scene changes and which commands accompany that change. DreamZero adds action and state interfaces to a pretrained video DiT, allowing both noisy video latents and noisy action vectors to influence the current chunk. Equation (1) offers an inverse-dynamics interpretation: a visual future helps specify the actions needed to reach it. The implementation does not first finish an independent video plan and then call a separate IDM. Instead, Equation (3) trains joint velocity prediction, and Figure 14 exposes the current modalities to each other while restricting history to visual conditioning. This distinction explains why better action generation can depend on video representations even when the final product delivered to the robot is a motor command. e03e04e05
Figure 5. Flash emphasizes noisy video conditioning while actions keep a uniform training schedule. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the arrows beneath the axes: small timestep means high noise, because Equation (2) interpolates from Gaussian noise at zero to clean data at one. The left red curve therefore emphasizes unreliable visual latents; it does not emphasize nearly clean video. Its legend gives the transformation t = 1 − η with η sampled from Beta(7,1), consistent with Equation (5). The right curves overlap because action timesteps remain uniform. The standard model shares one timestep across modalities; Flash decouples them. Read these curves as training distributions that expose action prediction to noisy visual context, not as measured robot-performance curves or an inference trajectory. e04e07e08e16
What it supports. The intended schedule has expected video timestep 0.125, compared with 0.5 under uniform sampling. It targets the mismatch encountered when few-step inference must produce useful actions before the visual representation is fully refined. Table 3 provides the downstream test of whether this training change helps.
Where the evidence stops. Algorithm 1, line 13, prints t_vid ∼ Beta(7,1), omitting the 1 − η transformation and reversing the bias under Equation (2). This faithful figure agrees with the main text and Equation (5); the supplied PDF does not resolve the pseudocode discrepancy.
5.2 Use real feedback to stop imagined history from accumulating
An autoregressive generator normally conditions on its own earlier outputs, allowing visual mistakes to propagate. DreamZero has an additional source of information: the robot can observe the scene after acting. Figure 4 and Algorithm 2 therefore return real images to the KV cache and discard the predicted future from recurrent history. My interpretation is that this corrects the model’s account of the past without certifying its next plan. Appendix H makes the distinction concrete: execution can follow a mistaken video prediction. Asynchronous control also keeps the robot moving while the next chunk is computed. A 1.6-second action horizon does not mean waiting 1.6 seconds between policy updates, and the advertised approximately 7 Hz is a chunk-generation rate, distinct from the AgiBot controller’s 30 Hz action rate. e03e05e06e18e20
5.3 Track what the policy has seen before calling it zero-shot
The evaluation changes the meaning of prior exposure across experiments. Figure 8 tests familiar robot tasks in unfamiliar environments and with unfamiliar objects. Figure 9 adds tasks absent from the embodiment’s robot training, although the pretrained video model has broader prior experience. For DROID, the split is explicitly based on verbs, and the toaster-lever example shows why lexical novelty is an imperfect proxy for motion novelty. Table 2 then supplies demonstrations of nine target tasks through another robot or humans, so its improvement is transfer after visual exposure. Finally, the YAM adaptation experiment adds about 30 minutes of target-embodiment data and reports qualitative novel-object examples. My reading is that these results support several distinct transfer abilities; they should not be pooled into one universal zero-shot success claim. e09e10e11e12e14e15e21
5.4 Training and inference
During training
Train AgiBot and DROID models separately for 100K steps, global batch 128. Update all DiT blocks and state/action modules; freeze the text encoder, image encoder and VAE. Remove idle actions and use relative joint positions. AgiBot contributes about 500 hours across 22 environments. e09e10
Post-train 50K steps per task using 33 hours of shirt folding, 12 of fruit packing, or 40 of table bussing. Flash is mainly a final training stage: video timesteps favor high noise while action timesteps remain uniform. e13e07
During inference
Prefill visual KV history, jointly denoise a fresh chunk, filter actions, and execute asynchronously. Real images and state refresh conditioning; predicted video is discarded from subsequent history. This avoids recycling hallucinated frames but cannot guarantee correct current actions. e06e20
CFG runs across two GPUs; velocity caching skips DiT evaluations when successive predictions align sufficiently. Compilation, GPU scheduler operations and mixed precision further reduce latency. Approximately 7 Hz denotes chunk generation on two GB200s; AgiBot motor control remains 30 Hz. e19e18e06
Action smoothing uses cubic 2× upsampling, a Savitzky–Golay filter with window 21 and polynomial order 3, then downsampling. e07
5.5 Implementation flow
- Encode and couple modalities
Wan2.1-I2V-14B-480P supplies the backbone. A VAE encodes observations; added state/action encoders and an action decoder connect robot signals to the shared DiT. Equation (1) interprets the joint distribution as video prediction times an implicit IDM, although deployment uses one joint model. e03e09
- Train with clean visual history
Teacher forcing denoises each noisy video/action chunk using previous clean context. Equation (3) fits joint clean-minus-noise velocity by weighted squared error. Figure 14 permits current video/action interaction and attention to earlier conditioning frames; previous action predictions are not recurrent history. e04e05
- Match time across modalities
Each chunk contains two latent video frames and spans 1.6 seconds: 48 actions at 30 Hz for AgiBot, or 24 at 15 Hz for DROID. Video is sampled at 5 FPS. Four chunks give 33 raw frames, reported as 6.6 seconds of context. e05
6. Experiments & results
DreamZero turns a pretrained video diffusion transformer into a policy that jointly denoises visual futures and motor commands, then refreshes its history with real observations. Real-robot results support stronger generalization than the tested VLAs, but single-step acceleration costs some task progress and uses substantial hardware. “Zero-shot” refers to held-out robot tasks or environments, not learning without robot data.
6.1 Read the original evidence
Figure 9. Unseen-task execution improves, but progress and complete success tell different stories. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend to distinguish model families and initialization. “Scratch” here excludes prior robot pretraining; DreamZero still starts from a pretrained video generator. The upper row and lower-left groups show individual AgiBot tasks, while the lower-right groups summarize two different embodiments. Read the axis labels carefully: AgiBot’s aggregate is task progress, and DROID separately reports progress and success. Compare methods within each group before comparing embodiments, whose task sets and trial counts differ. The AgiBot protocol uses eight rollouts per task; DROID uses two. Variation across tasks is substantial, so the aggregate is not evidence of uniformly reliable novel-skill execution. e09e10e12e21
What it supports. DreamZero reaches 39.5% AgiBot progress versus 16.3% for pretrained π0.5. On DROID it reaches 49% progress but only 22.5% success. The AgiBot task bars range from 8.2% for map folding to 85.7% for hat removal, showing that the generalization gain remains strongly task-dependent.
Where the evidence stops. Figure 9 does not define its error bars. DROID novelty is based on absent verbs: Appendix G includes seen “push” and unseen “depress” for a toaster lever. This weakens an interpretation that every DROID test requires an entirely new physical motion.
Table 2. Small amounts of task-specific donor video improve the target robot’s progress. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the first row as the no-transfer DreamZero checkpoint, then compare each donor-video condition against that row. The numbers are average task progress; the original caption defines the ± values as standard errors. Both donor datasets contain 72 multi-view trajectories covering nine target tasks. Human videos total 12 minutes and YAM videos 20 minutes. Training continues for 10K updates using a 1:1 mixture with AgiBot data: donor examples supply only the video objective, while AgiBot examples retain joint video-action supervision. These are tasks initially absent from target-robot training, but the transfer models have now seen demonstrations of them in another embodiment. e14e12
What it supports. Human video increases progress from 38.3% to 54.3%, and YAM video increases it to 55.4%: gains of 16.0 and 17.1 percentage points. This supports transfer through visual supervision despite absent donor action labels. It does not demonstrate learning a new target robot without prior AgiBot action data.
Where the evidence stops. The nine-task baseline excludes pulling cart and must not be substituted for Figure 9’s ten-task mean. The caption points to a nonexistent Table 9 in this PDF. No significance test or continued-training-only control is reported alongside the transfer rows.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| AgiBot unseen-task generalization 10 held-out tasks, unseen environments and objects; eight rollouts per task across four robots (80 per checkpoint). | 39.5% Average task progress (%) | Pretrained π0.5: 16.3%; pretrained GR00T N1.6: 5%. A 23.2-percentage-point gain over the strongest listed VLA. This is partial completion, not full success. e10e12 |
| AgiBot seen tasks in new environments 10 tasks present in robot pretraining; eight rollouts per task, unseen environments and objects. | 62.2% Average task progress (%) | Pretrained π0.5: 27.4%; pretrained GR00T N1.6: 8.4%. Supports environment/object generalization; it does not test unseen task motions. e10e11 |
| DROID unseen-verb evaluation 20 tasks whose verbs are absent from DROID; two rollouts per task, object positions matched across checkpoints. | 49% progress; 22.5% success Average task progress; full success (%) | GR00T N1.6: 31% / 12.5%; π0.5: 33% / 7.5%. Execution improves, but most trials do not fully succeed. Verb novelty does not ensure motion novelty. e10e12e21 |
| AgiBot task-specific post-training Shirt folding, fruit packing and table bussing in unseen environments; 10 rollouts per task. | 90.5% average; 92.5%, 96%, 83% respectively Average task progress (%) | Pretrained π0.5: 79.8% average; 92.5%, 71%, 76%. A 10.7-point average improvement after task-specific training. e13 |
| Video-only transfer to AgiBot Nine previously unseen tasks, excluding pulling cart; 72 multi-view trajectories per donor, 1:1 mix with AgiBot data for 10K updates. | Human: 54.3% ± 10.4%; YAM: 55.4% ± 9.5% Average task progress ± standard error (%) | Without transfer: 38.3% ± 7.6%; human videos total 12 minutes, YAM 20 minutes. Gains are 16.0 and 17.1 points. Target-task videos are now supplied; this differs from the untouched 10-task test. e14 |
| Single-step table bussing with DreamZero-Flash Table 3 compares four-step inference, naive one-step inference and one-step Flash. | Flash: 74% ± 10.1% at 150 ms Task progress ± standard error (%); inference latency | Four-step: 83% ± 6.1% at 350 ms; naive one-step: 52% ± 10.2% at 150 ms. Flash recovers 22 points over naive one-step, but trails four-step by 9 points at 2.33× lower latency. No significance test is supplied. e16 |
| PnP Easy model and data ablations 50K updates, batch 32; equal 500-hour data budgets for diversity comparison. | Diverse 14B AR: 50% ± 6.3% Task progress ± standard error (%) | Repetitive 14B AR: 33% ± 4.2%; diverse 5B AR: 21% ± 4.2%; diverse 14B BD: 50% ± 14.4%. Diversity and scale help this restricted test. AR and BD have equal mean progress; smoother AR motion and faster inference are separate claims. e17 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Flash recovers much of the one-step performance loss, with a remaining accuracy cost. Original paper, p. 17 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the first two rows to isolate the effect of reducing denoising steps without Flash training: latency falls, but progress also drops. Then compare the second and third rows, which share one-step inference and 150 ms latency, to see what the altered training schedule recovers. Finally, compare the first and third rows to assess the actual deployment tradeoff. The ± values are standard errors according to the caption. The last column is relative to the four-step row, so its 2.33× speedup is different from Table 1’s cumulative 38× figure, which includes multiple system and implementation changes. e16e19
What it supports. Naive one-step inference falls from 83% to 52% progress; Flash raises the one-step result to 74%. It recovers 22 percentage points while remaining 9 points below the four-step model. Thus the table supports a useful speed–accuracy tradeoff, rather than exact preservation of the baseline score.
Where the evidence stops. The table evaluates table bussing. It does not establish the same tradeoff across unseen tasks or all embodiments. Nearby prose calls 74% a success rate, but the table header and caption identify task progress; this edition follows the table.
Table 4. Diversity and model size help; autoregression does not increase the reported mean. Original paper, p. 17 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the three blocks as separate comparisons. In Q1, model size and architecture stay at 14B autoregressive while the data distribution changes. The text specifies equal 500-hour budgets, with the repetitive set spanning 70 tasks. In Q2, compare 5B and 14B DreamZero on diverse data; the VLA rows are additional size-matched baselines, not the same pretrained backbone with a single loss removed. In Q3, compare bidirectional (BD) and autoregressive (AR) DreamZero at 14B. Every row uses 50K training steps and batch 32, evaluated on PnP Easy; these are smaller training runs than the main models. e17e18e09
What it supports. Diverse data raises the 14B AR mean from 33% to 50%, and scaling from 5B to 14B raises it from 21% to 50%. AR and BD both score 50%, with reported standard errors of 6.3% and 14.4%. The table therefore does not demonstrate a mean-progress advantage from autoregression.
Where the evidence stops. The ± terms are standard errors, not measures of motion smoothness. Smoother AR execution and 3–4× faster inference are separate prose claims. Equal update counts and data hours do not establish equal compute, nor does this restricted benchmark establish a general scaling law.
7. Analysis & limitations
7.1 What the evidence leaves open
Authors identify expensive inference, short context, limited scaling evidence, small in-lab human-video transfer, and difficulty with sub-centimeter precision. Tasks requiring memory were explicitly not evaluated. e18e06
“Scratch” still uses foundation-model initialization. Matching batch size and updates does not establish equal FLOPs or pretraining experience. Figure 9 does not define its error bars; Tables 2–4 identify standard errors without explaining aggregation. e10e12e14e16e17
YAM adaptation uses 55 trajectories across 11 tasks, about 30 minutes, but gives qualitative examples rather than aggregate success. Two failures align bad video plans with bad execution; they cannot establish the frequency or cause of all failures. e15e20
7.2 Questions for discussion
- Would Flash recover single-step performance on unseen contact-rich tasks beyond table bussing?
- How much DROID unseen-verb performance survives a split excluding synonymous motions?
8. Reproducibility audit
8.1 Requirements and known gaps
Replication needs Wan initialization, embodiment data, normalization, idle filtering, synchronized observations and the Figure 14 mask. The PDF omits optimizer/learning-rate settings, training hardware, numerical idle/cache thresholds and detailed per-task scoring rubrics. Appendix G provides prompts/scenes, not those rubrics. e09e05e19e21
Reproduce the two-GB200 configuration and compiled mixed-precision pipeline for speed; Appendix D specifies PyTorch ≥2.9 for cuDNN SDPA. Separate warm-up from steady-state latency and preserve smoothing settings. e18e19e07
Resolve Flash sampling before training. Algorithm 1 includes past actions in context although Figure 14 masks them out; Algorithm 2 labels clean cache inputs t=0 although Equation (2) uses t=1 for clean interpolation. Pseudocode alone is incomplete. e08e04e05e06
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: Test the Flash schedule in both printed orientations
Reader-proposed check, not an experiment performed here: start three runs from the same trained DreamZero checkpoint. Continue training with coupled uniform timesteps, with Equation (5)’s 1 − Beta(7,1) video schedule, or with Algorithm 1’s direct Beta(7,1) schedule. Match examples, updates, action timesteps and seeds. Verify sampled video-time means before training, then compare one-step task progress and latency on matched table-bussing scenes using identical hardware and smoothing; retain a four-step reference. The proposed mechanism predicts better one-step recovery under noisy-video-biased training. Equal or superior recovery with the opposite bias would weaken that account and would leave the paper’s implementation choice unresolved. e04e07e08e13e16e19
Check 2: Separate donor-task information from extra training
Reader-proposed check, not an experiment performed here: continue the same AgiBot checkpoint for the reported 10K updates under three matched conditions: AgiBot-only continuation, a 1:1 mix with unrelated donor videos, and the reported 1:1 mix with donor videos of the nine target tasks. Match donor duration, view format, total updates and evaluation scenes; mask donor action loss in both video conditions. Score the nine transfer tasks separately from untouched tasks, using explicit partial-credit and full-success rubrics with repeated trials. A target-video-specific gain would support transfer of task information; similar gains from AgiBot-only continuation or unrelated video would weaken attribution to cross-embodiment task demonstrations. e10e14e21
8.3 Reading coverage
Visual audit: Visually inspected the title/author page and PDF pages 2–29, including every figure (1–16), quantitative Tables 1–4, evaluation setups in Tables 5–7, the attention masks, both algorithms, and implementation details. All six final original crops were viewed; architecture and noise-schedule bounds were corrected and re-viewed. Figure 5 agrees with Equations (2) and (5), but Algorithm 1 omits the timestep complement; this is disclosed. Table 3’s task-progress label is retained despite conflicting nearby prose. Reference pages 30–36 were read in the text chunks only. External videos, code and separate supplements were outside this inspection.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title and author page (p. 1)
- Abstract and Introduction (pp. 2–4)
- Related Work (pp. 4–5)
- DreamZero: architecture, objectives and real-time execution (pp. 6–10)
- Experimental Setup (pp. 10–13)
- Experimental Results and ablations (pp. 13–18)
- Discussion and Acknowledgment (pp. 18–19)
- Appendices A–B: alternatives and autoregression (pp. 20–21)
- Appendix C: attention masks, configuration and Algorithms 1–2 (pp. 21–22)
- Appendix D: execution and Flash details (pp. 22–24)
- Appendix E: data collection (pp. 24–25)
- Appendices F–G: evaluation setups (pp. 25–28)
- Appendix H: failure cases (pp. 28–29)
- References (pp. 30–36)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction limitation was addressed by inspecting PDF pages 1–29 and all six final crops. All 11 text chunks were read individually, including references; reference pages 30–36 were read as text only.
- Identity/version note: the observed title and author list match the catalog. This artifact is arXiv:2602.15922v1, stamped 17 February 2026, while its title-page header says 2026-2-19. No other revision or edition was supplied or compared. The catalog BibTeX expands Ayaan Malik and Nadun Ranawaka, omits Jim from Linxi Fan, and orders Ryan Julian differently; verified metadata follows the PDF title page. The catalog workshop venue is not established by this PDF.
- Code, model weights, datasets, project videos and separate supplements were not inspected; no experiment was reproduced. Release statements are author statements only.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author block, affiliation and arXiv margin
Title is World Action Models are Zero-shot Policies; authors are affiliated with NVIDIA. Margin identifies arXiv:2602.15922v1, 17 Feb 2026; header says 2026-2-19.
Go to primary source ↓e02PDF p. 2, Abstract and §1
Motivates motion generalization using video priors and heterogeneous robot demonstrations; introduces DreamZero as a 14B WAM.
Go to primary source ↓e03PDF pp. 6–7, Figure 4 and §3.1, Equation (1)
Single shared DiT jointly predicts video and action from visual history, language and proprioception; Equation (1) gives the video/IDM factorization. Multiple views are concatenated.
Go to primary source ↓e04PDF pp. 7–8, §3.1, Equations (2)–(3)
Teacher forcing and shared within-chunk timesteps; interpolation runs from noise at 0 to clean data at 1. Joint velocity target is clean minus noise with weighted squared-error training.
Go to primary source ↓e05PDF p. 21, Appendix C, Figure 14 and configuration paragraph
Attention uses earlier clean visual conditioning and current video/action blocks, excluding earlier action blocks. Two latent frames per chunk, four chunks, 5 FPS video, 48/24 actions at 30/15 Hz, and 6.6-second context are specified.
Go to primary source ↓e06PDF p. 8, Model Inference, §3.2.1–3.2.3 and footnote 2; p. 22, Algorithm 2
Asynchronous execution refreshes KV history with real observations and discards predicted video. Baseline latency is 5.7 seconds. Memory-dependent tasks were not evaluated. Algorithm 2 calls clean cache conditioning t=0.
Go to primary source ↓e07PDF p. 9, Figure 5 and §3.2.5; p. 24, Appendix D.3, Equations (5)–(6) and Action Chunk Smoothing
Flash video time is 1 minus a Beta variable; action time remains uniform. Beta(7,1) yields mean video timestep 0.125. Filtering uses cubic 2× upsampling and Savitzky–Golay window 21/order 3.
Go to primary source ↓e08PDF p. 22, Algorithm 1 lines 9, 13, 19–22; pp. 7, 9, 21, 24, Equations (2), (5) and Figures 5, 14
Algorithm 1 directly samples video time from Beta(7,1), conflicting with Figure 5 and Equation (5). Its context includes past actions, whereas Figure 14 shows no attention to previous action blocks.
Go to primary source ↓e09PDF pp. 10–11, §4.1 and footnotes 5–7
About 500 AgiBot hours from 22 environments; Wan2.1-I2V-14B-480P; separate 100K-step, batch-128 training for AgiBot and DROID; frozen text/image encoders and VAE, idle filtering and relative joints. AgiBot dataset release is planned.
Go to primary source ↓e10PDF pp. 10–12, §4, §4.1 Evaluation Protocol; p. 25, Appendices F–G
Scratch means foundation weights without prior robot training; pretrained baselines use robot-trained checkpoints. AgiBot tests 10 seen plus 10 unseen tasks, eight rollouts per task; DROID tests 20 plus 20, two each. Evaluation environments and objects differ from training.
Go to primary source ↓e11PDF p. 13, Figure 8, AgiBot AVG (Task Progress), and §5.1 Q1
Seen-task AgiBot average progress: DreamZero 62.2%, pretrained π0.5 27.4%, pretrained GR00T N1.6 8.4%.
Go to primary source ↓e12PDF p. 14, Figure 9, AVG panels and §5.1 Q2
AgiBot unseen progress: 39.5%, 16.3%, 5% for DreamZero, pretrained π0.5, pretrained GR00T. DROID progress/success: 49/22.5, 33/7.5, 31/12.5. DreamZero AgiBot task scores include hat removal 85.7% and map folding 8.2%.
Go to primary source ↓e13PDF pp. 12–13, §4.2; p. 15, Figure 10 and §5.1 Q3
Post-training data: folding 33 h, packing 12 h, bussing 40 h; 50K updates per task and 10 evaluation rollouts each. DreamZero scores 92.5/96/83 (mean 90.5); pretrained π0.5 scores 92.5/71/76 (mean 79.8).
Go to primary source ↓e14PDF pp. 15–16, §5.1 Q4, Figure 11, Table 2 and footnote 10
Donor data uses video-only loss, mixed 1:1 with joint-loss AgiBot data for 10K updates. Nine tasks exclude pulling cart. Human/YAM data are 72 trajectories and 12/20 minutes. Progress ± SE: 38.3±7.6 baseline, 54.3±10.4 human, 55.4±9.5 robot. Caption refers to an absent Table 9.
Go to primary source ↓e15PDF pp. 16–17, §5.1 Q5, Figure 12 and footnotes 11–12
AgiBot checkpoint adapts to YAM with 55 trajectories over 11 tasks (about 30 minutes). Novel-object language following is illustrated; no aggregate adaptation score is supplied.
Go to primary source ↓e16PDF p. 17, §5.1 Q6 and Table 3, all rows and caption
Bussing progress ± SE: four-step 83±6.1% at 350 ms; one-step 52±10.2% at 150 ms; Flash one-step 74±10.1% at 150 ms. Table says task progress, despite nearby prose saying success rate.
Go to primary source ↓e17PDF pp. 17–18, Table 4 and §5.2 Q1–Q3
Ablations use 50K steps/batch 32 on PnP Easy. Diverse/repetitive data each total 500 h. Diverse 14B AR: 50±6.3%; repetitive: 33±4.2%; 5B: 21±4.2%; 14B BD: 50±14.4%. AR smoothness and 3–4× speed are separate prose claims.
Go to primary source ↓e18PDF pp. 18–19, §6
Authors acknowledge limited scaling evidence, only 12 minutes of in-lab human data, two-GB200 inference, short visual memory, and difficulty with sub-centimeter precision.
Go to primary source ↓e19PDF pp. 8–10, §3.2 and Table 1; pp. 22–23, Appendix D.1–D.2
Optimizations include CFG parallelism, velocity caching, compilation, cuDNN attention, GPU scheduler and NVFP4 mixed precision. Table 1 reports 38× total speedup; Appendix D requires PyTorch ≥2.9 for stated SDPA backend and notes first-trajectory recompilation. Cache threshold is unspecified.
Go to primary source ↓e20PDF pp. 28–29, Appendix H and Figure 16
Two failures pair generated and executed behavior: passing a marker instead of drawing, and grasping bread before opening the oven. They illustrate alignment with unsuccessful visual plans.
Go to primary source ↓e21PDF p. 12, DROID Evaluation Protocol; p. 25, Appendix G; p. 28, Table 7(a–b)
DROID novelty is defined by absent verbs. Seen tasks include pushing a toaster lever down; unseen tasks include depressing the lever. Appendix G gives scenes/prompts, without detailed partial-credit rubrics.
Go to primary source ↓8.5 Primary sources
World Action Models are Zero-shot Policies ↗
PDF · 17,822 extracted words
Source fingerprint
cb91d24a04717a6c77fd39594af7f3f85a05317b7078b0d9c8d8b4ddd932f483