Riemann-1.0: An Embodied World Action Model for Physical AI
1. Paper overview
In one sentence: Action-first autoregression lets one transformer serve as a robot policy and a visual simulator, while the reported evaluations leave the contribution of its ordering and staged supervision unresolved. E03E04E07E08E09E11E13E16E17
| At a glance | What to know |
|---|---|
| Research problem | Author claim How can heterogeneous human videos, handheld demonstrations, and robot trajectories train executable control and visual dynamics together? The sources differ in control semantics, temporal density, and label reliability. The authors propose a common trajectory interface and causal action-before-observation modeling. E02E03 |
| Core mechanism | |
| A key reported result | Four real-world manipulation tasks: Reported SR 85.00%; PSR 94.43%. Average final-task SR and milestone-based PSR (%). Tianji Marvin dual-arm robot; joint post-training on ordered cube stacking, kitchen organization, clothes folding, and desk organization. G0.5: 70.00% SR, 80.18% PSR; SR gain 15.00 percentage points. Physical execution results. Main-suite trial counts and uncertainty are absent. E10E11E17 |
| Reading caution | Reader analysis No controlled ablation isolates causal ordering, visual loss, LAM, curriculum stages, or balancing. Architecture, data, and schedule effects remain confounded. E17 |
Core contributions
Figure 4. One temporal backbone, with control interfaces selected for each embodiment. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start on the left: packed camera views pass through the Wan VAE and patch embedding, while the structured task prompt passes through T5. The embodiment ID selects the control interface rather than a separate transformer. Follow the unified trajectory tokens into the central Action/Video DiT, then trace the two right-hand branches. The upper branch predicts visual-latent velocity; the lower predicts action velocity in the selected robot’s representation. Their masked losses combine at the bottom. Read this diagram together with Equation (1): current actions condition visual consequences, whereas action prediction uses the preceding interaction history. The shared backbone does not erase embodiment-specific action semantics. E03E04E05E06E07E17
What it supports. The figure directly supports the One Model architecture assessment: visual and action prediction share the DiT, with different output interfaces. It also explains how visual prediction can remain a training signal when deployed robot control requests actions and obtains subsequent observations from the environment.
Where the evidence stops. The two branches do not imply that both must run before every robot action. Robot states are conditioning inputs, not generated outputs. The diagram supplies no parameter count, latency, or evidence that the architecture alone causes the reported gains.
2. Motivation
2.1 The problem and the proposed response
How can heterogeneous human videos, handheld demonstrations, and robot trajectories train executable control and visual dynamics together? The sources differ in control semantics, temporal density, and label reliability. The authors propose a common trajectory interface and causal action-before-observation modeling. E02E03
2.2 What this reading follows
Riemann-1.0 makes the action chunk the connection between robot control and visual prediction. A shared transformer learns both, but deployment can choose which output to request: real robots execute an action and return camera observations, while the visual simulator predicts those observations from supplied actions. Its training curriculum starts with motion codes inferred from human video, then introduces structured trajectories and finally robot-only demonstrations. The reading below follows that information flow before examining physical success rates and compositional simulation results. The central evidence boundary is causal attribution: the paper compares complete systems, but does not isolate which architectural or curriculum choice produces the gains. E03E04E07E08E09E11E13E16E17
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 | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
One Model is supported by the shared transformer, despite specialized heads and a separate training-time LAM. Other mechanisms/Outside quadrants fits action-first autoregression: it does not extract actions by inverse dynamics from imagined futures or require simultaneous action–future denoising. Alignment follows from staged supervision; memory support is narrower, through cached history and long tasks without a memory ablation. E03E04E07E08E09E11E17
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 1. Put the action between two observations
Begin with the factorization rather than the word world model. At each transition, Riemann first predicts an action chunk from the preceding observation, state, and action history. Its visual prediction factor receives that same history plus the current action. This ordering creates two operating modes with the same interfaces. In robot deployment, the action is executed and the environment supplies the next camera views and state; the model need not generate an imagined future to choose the action. In visual simulation, the latent head supplies the next visual observation instead. Training uses clean history and noisy flow-matching targets, with structured attention excluding unavailable future clean observations. Local attention within a generation block remains allowed. Thus the causal claim concerns transition dependencies, not an assertion that every individual token is processed strictly one at a time. E03E05E06E07E16
Figure 7. The supervision changes meaning while the action-video training interface stays consistent. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the three cards from left to right, paying attention to both data and loss weight. Stage I obtains latent pseudo actions from a separately trained, frozen LAM; those codes explain visual transitions but are not robot commands. Stage II replaces them with UMI, robot, and reconstructed human-hand trajectories. Embodiment-specific interfaces retain the meaning of each action space. Stage III removes human and UMI sources and concentrates on robot demonstrations. The λ values at the bottom weight the action objective: 0.1, 0.5, then 0.9. Section 4.2 on PDF page 14 raises this to 0.95 for real-world post-training, outside the three-stage diagram. E08E09E10E17
What it supports. The curriculum provides a concrete route from abundant video to executable control: first learn motion associations, then align trajectory semantics, then specialize robot actions. Visual supervision remains active even in the action-focused stages. This is a specified training recipe, while its claimed performance advantage remains a hypothesis about the complete system.
Where the evidence stops. Changing stages changes both the source mixture and the loss weight. No ablation separates those effects or isolates the LAM contribution. The paper also leaves the sampling weights and optimization budgets needed to reproduce the transitions unspecified.
5.2 2. Understand what a pseudo action can supervise
A human video does not contain the robot command that would reproduce its motion. Stage I therefore uses a latent-action VAE as an annotator: a compact code must help reconstruct the next frame from the current one. Once trained, the annotator is frozen and its posterior mean becomes deterministic supervision. These codes fit the WAM temporal interface, but their meaning is initially a visual transition rather than executable control. Stage II introduces trajectory labels with physical or hand-pose semantics; Stage III keeps only robot demonstrations. My interpretation is that this curriculum preserves the computational role of the action stream while progressively changing what its values mean. The increasing action-loss weight reinforces the transition toward control. The source describes this rationale, but supplies no controlled experiment proving that the ordering of these stages is necessary. E08E09E04E17
5.3 3. Separate task success from model fidelity
The experiments answer three different questions. Table 1 measures whether the deployed robot completes four post-trained tasks and how much milestone progress it makes. Table 2 tests a small set of held-out instructions and objects. Tables 3–5 measure executed task success inside external simulation benchmarks. Figure 10 instead shows observations generated by Riemann itself. These should not be collapsed into one world-model score: successful policy execution does not measure the accuracy of predicted video, and a plausible video does not establish a successful action sequence. My reading is that the strongest evidence concerns the complete policy system, especially kitchen organization and RoboCasa’s composite groups. The claimed simulator role is architecturally concrete but empirically less characterized. Its action sensitivity, long-rollout error, and usefulness for selecting plans remain open evaluation questions. E11E12E13E14E15E16E17
5.4 Training and inference
During training
Stage I trains then freezes a latent-action VAE. Adjacent-frame transitions pass through a 32-dimensional bottleneck under reconstruction and weak KL losses; deterministic posterior means supply pseudo actions. With λ=0.1, WAM learns dynamics associations, not a deployable policy. E08
Stage II uses UMI, robot, and reconstructed human-hand trajectories with λ=0.5. Stage III retains only high-quality robot demonstrations with λ=0.9. Visual supervision remains active throughout. E09
Post-training jointly fits one generalist to four real-world tasks with λ=0.95. Section 4.2 gives three teleoperation hours per task; Section 5.1.1 gives 15 demonstrations per task without reconciling the descriptions. E10
During inference
Initialize a KV cache with prompt, observation latent, and state. Denoise Gaussian action noise, cache and execute the chunk for apf low-level steps, then append real observations and final state. This policy loop does not require future-video generation. E07
Simulator mode conditions the latent head on policy, candidate-plan, or recorded actions and feeds predicted visuals back into history. States remain observed or replayed rather than generated, limiting its interpretation as an autonomous physical-state simulator. E03E16
When the fixed cache window fills, history resets around the latest observation. Window length and reset effects are unreported. Simulation benchmarks use 16 low-level action steps per visual latent. E07E13E17
5.5 Implementation flow
- Standardize embodied experience
Section 2.1 reports 230K+ hours overall: 200K+ human, 12K+ handheld/wearable, and 20K+ robot hours. Semantic segmentation, MANO hand reconstruction, camera-pose estimation, calibration, filtering, and semantic-aware sampling produce aligned trajectories. Embodiment-specific normalization preserves distinct physical control spaces. E02
- Encode observations and control
T5 encodes task, embodiment, and camera-view prompts. The Wan VAE and 3D patch embedding encode a packed multi-view canvas. Embodiment IDs select state/action projections and action heads around a shared transformer; visual prediction uses a shared latent head. E04
- Prevent unavailable future information
Teacher-forced clean history and noisy flow targets occupy one sequence. Structured visibility blocks future clean information while allowing attention within local generation blocks. Action chunks align to visual transitions. Validity masks exclude padding and invalid targets; padded channels remain zero during noise injection. E03E05E06
6. Experiments & results
Riemann-1.0 learns action prediction and visual dynamics in one transformer, ordering actions before their consequences. Robot deployment uses actual camera feedback; a second operating mode generates action-conditioned video. Pretraining moves from human-video pseudo actions through mixed trajectories to robot-only supervision. Strong reported task results accompany qualitative simulator examples, but no controlled ablation separates the effects of architecture, curriculum, and data scale.
The supplied paper contains no controlled ablation table, ablation plot, or diagnostic isolating action ordering, LAM, curriculum stages, loss weights, or cache resets. Accordingly, this edition includes architecture and curriculum figures, three quantitative comparison tables, and qualitative simulator examples, but no ablation visual. Table 2 diagnoses a task-distribution boundary while changing several requirements together; it is not relabeled as an ablation. Figure 10 provides no quantitative fidelity or counterfactual-action evaluation. E12E16E17
6.1 Read the original evidence
Table 1. The largest physical-task separation is kitchen organization; other tasks include ties. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task occupies two columns: SR is final completion, while PSR measures progress through task-specific milestones. Begin with the rightmost averages, then inspect individual tasks to see what drives them. Riemann reaches 85.00% average SR against G0.5’s 70.00%, but kitchen organization is unusually influential: its SR is 90.0% against 35.0% for the best competitor. Clothes folding and desk organization instead tie G0.5 in SR at 85.0% and 80.0%. Keep the retained DreamZero* note in view: that row is the authors’ reproduced implementation trained on their manipulation data, not simply an official checkpoint evaluation. E10E11E17
What it supports. The table establishes a reported 15-percentage-point average SR advantage on this four-task physical setup. It also shows why average performance needs a task breakdown: kitchen storage provides a large separation, while folding and desk organization primarily improve milestone progress relative to the strongest SR comparator.
Where the evidence stops. The paper does not supply main-suite trial counts, uncertainty intervals, or a complete PSR scoring rubric. It describes a shared real-world setting without establishing matched pretraining data and compute for every baseline.
Table 2. Held-out instructions expose a drop when color, receptacle, and execution-order constraints accumulate. Original paper, p. 17 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the successful-trial fractions before the averages. Riemann completes eight of ten simple cube-placement trials and five of ten trials with added constraints; the 65.0% compositional average pools these two task scores. The lower block tests Rubik’s-cube storage and towel placement without extra demonstrations or parameter updates. Those outcomes are ten of ten and seven of ten, producing the 85.0% OOD average. The final 75.0% averages all four tasks and should not be confused with the post-trained suite’s 85.0% SR. Here OOD means absent from real-world post-training, not verified absent from the entire pretraining corpus. E12E02E17
What it supports. The results support limited transfer to new instructions and the two specified OOD tasks. The constrained compositional task succeeds only half the time even though it leads the listed baselines. That contrast is a useful diagnostic of the reported generalization boundary, without attributing failures to any single instruction component.
Where the evidence stops. Each task has only ten trials and no reported uncertainty. Several requirements change together, so this is not a controlled mechanism ablation. Figure 9 calls the Rubik’s-cube receptacle a basket; the table and prose call it a storage box.
Table 3. RoboCasa’s largest margins occur in the composite task groups. Original paper, p. 18 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down each column to compare systems under the reported RoboCasa protocol: pretraining on 300 tasks, then fine-tuning on Target-50 before evaluating the 50 target tasks. The bottom two rows make the strongest comparison. Riemann exceeds ABot-M0.5 by 3.6 points in Atomic-Seen, 11.7 in Composite-Seen, and 10.7 in Composite-Unseen. The source-reported average rises from 54.2% to 62.6%. Keep the group labels attached to every score; the word Unseen belongs to this benchmark’s protocol. The accompanying section does not justify reinterpreting the entire evaluation as deployment without downstream fine-tuning. E13E14E15E17
What it supports. The reported 8.4-percentage-point average improvement is substantially larger than the 0.2-point RoboTwin and 0.1-point LIBERO average leads. Within RoboCasa, composite groups account for the larger differences, supporting the relevance of multi-step task evaluation to the paper’s claims about transferring manipulation skills.
Where the evidence stops. These are task-completion scores in a benchmark simulator, not measures of Riemann’s generated-video fidelity. No uncertainty or matched-budget ordering/curriculum ablation is provided, so the source does not isolate the cause of the stronger composite scores.
Figure 10, first three embodiment rows. Selected generated rollouts illustrate the simulator interface across three embodiments. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each row is a separate example; read horizontally through the uniformly sampled generated frames. The left-hand labels identify the embodiment and abbreviated task, while the columns show successive visual states. The first and third rows involve cloth; the middle row shows a carrot-picking scene. This crop contains the header and first three complete rows of the ten-row original figure. Section 5.3 explains the hidden conditioning: a current observation, prompt, robot state, and supplied action chunk enter the model, and its latent head predicts future camera observations. The actual action values and corresponding recorded ground-truth futures are not plotted. E18E16E03
What it supports. These examples show what the visual-simulator output looks like across different robot appearances and camera scenes. They accompany the authors’ claim of action-consistent visual dynamics, but the crop alone cannot distinguish faithful action-conditioned prediction from a visually plausible continuation of the task.
Where the evidence stops. No counterfactual action pair, ground-truth comparison, error curve, or planning outcome accompanies these sequences. Future robot states are externally observed or replayed, leaving an additional dependency when interpreting the model as a simulator.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Four real-world manipulation tasks Tianji Marvin dual-arm robot; joint post-training on ordered cube stacking, kitchen organization, clothes folding, and desk organization. | Reported SR 85.00%; PSR 94.43%. Average final-task SR and milestone-based PSR (%) | G0.5: 70.00% SR, 80.18% PSR; SR gain 15.00 percentage points. Physical execution results. Main-suite trial counts and uncertainty are absent. E10E11E17 |
| Kitchen organization Post-trained physical utensil storage, plate placement, and bowl stacking. | Reported SR 90.0%; PSR 98.4%. SR and PSR (%) | Best competing SR: G0.5, 35.0%; best competing PSR: LingBot-VLA, 80.2%. The strongest comparator differs by metric; this large advantage is task-specific. E11 |
| Real-world compositional generalization Seen object categories, held-out instructions; 10 robot trials per task. | Placement 8/10; added color/order constraints 5/10; average 65.0%. Task counts and average SR | π0.5: 45.0%; LingBot-VA: 15.0% average SR. Transfer degrades with stricter instructions; four added requirements change together. E12 |
| Real-world out-of-domain manipulation Scenes, objects, and goals absent from real-world post-training; no additional demonstrations or updates; 10 trials per task. | Rubik’s cube to storage box 10/10; towel to basin 7/10; average 85.0%. Task counts and average SR | π0.5: 50.0%; LingBot-VA: 30.0% average SR. Exclusion from the much larger pretraining corpus is not established. E12E02 |
| RoboCasa365 300-task pretraining followed by Target-50 fine-tuning; evaluation on 50 target tasks. | Atomic-Seen 74.2; Composite-Seen 56.0; Composite-Unseen 56.3; reported average 62.6. Simulation task SR (%) | ABot-M0.5: 70.6/44.3/45.6; average 54.2. Average gain: 8.4 percentage points. Largest gains concern composite groups within this adaptation protocol; not an independently established zero-shot evaluation. E13 |
| RoboTwin 2.0 50 bimanual tasks; 50 clean plus 500 randomized training demonstrations per task. | Reported Clean 94.6; Randomized 94.0; average 94.3. Clean, Randomized, and average simulation SR (%) | ABot-M0.5: 94.0/94.2/94.1; average gain 0.2 percentage points. The average lead coexists with a 0.2-point deficit under randomization; uncertainty is unreported. E14E17 |
| LIBERO Spatial, Object, Goal, and Long suites; 10 tasks per suite, 50 demonstrations per task. | Reported 99.6/100.0/97.6/98.6; average 99.0. Simulation task SR (%) | Being-H0.5 and G0.5 each average 98.9%. The 0.1-point average lead lacks uncertainty estimates; Riemann does not lead every suite. E15E17 |
6.3 Ablations and diagnostic examples
No ablation or diagnostic visual is included in this edition. The source-specific visual limitations are recorded above.
7. Analysis & limitations
7.1 What the evidence leaves open
No controlled ablation isolates causal ordering, visual loss, LAM, curriculum stages, or balancing. Architecture, data, and schedule effects remain confounded. E17
Figure 10 supplies qualitative simulator examples, without prediction-error metrics, counterfactual action tests, or measured planning benefits. Generated images do not establish physical execution success. E16E18
PSR is defined through intermediate milestones, but the scoring rubric is not reconstructable. Main-task trial counts and uncertainty are absent; held-out tasks have ten trials each. E11E12E17
Matched pretraining data and compute across baselines are not established. DreamZero* is the authors’ 5B reproduction, not an untouched official checkpoint or Riemann’s parameter count. E11E17
7.2 Questions for discussion
- Does visual supervision improve executed policy success without test-time video generation? [E06, E07]
- How should future states be supplied for genuinely counterfactual visual simulation? [E03, E16]
- Which long-horizon tasks require history lost during cache resets? [E07, E11]
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction requires the curated data manifest, semantic sampling weights, embodiment mappings/statistics, camera packing, temporal alignment, validity masks, and source-specific filtering settings. The paper describes their roles but not a fully executable recipe. E02E04E05
Model size, optimizer, batch size, training steps/compute, denoising budget, latency, cache length, and LAM KL coefficient remain unspecified. The paper identifies a frozen LAM annotator, but does not establish T5/Wan VAE freezing or detailed initialization. E04E07E08E17
Reader-proposed checks: compare matched-budget ordering variants on executed SR and latency; test visual predictions against changed-action replay controls. The illustrated edition specifies controls and falsifiable observations. E03E07E16E17
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: Match resources before testing action-first ordering
Reader-proposed experiment, not performed: train action-first and joint action-video variants on the same fixed RoboTwin task subset, with identical shared-backbone capacity, initial weights, demonstrations, action interface, optimization budget, and loss weights. Predefine a clean/randomized test split and run multiple seeds with identical episode budgets. Compare executed SR, action-generation latency on the same hardware, and the number of visual denoising calls before an action. Also report an accuracy-versus-latency sweep rather than selecting one favorable denoising budget. A repeatable advantage at matched resources would support an ordering-specific contribution; similar performance would weaken attribution of the paper’s full-system gains to causal ordering alone. E03E04E07E14E17
Check 2: Test whether changed actions change the predicted world correctly
Reader-proposed experiment, not performed: create paired simulator episodes from the same observation, prompt, and initial robot state but execute two deliberately different valid action chunks. Record each actual visual future and state sequence. Generate Riemann rollouts for both actions using a fixed sampling seed, alongside shuffled-action and zero-action controls. Evaluate first-step object displacement and gripper position against recorded futures, then measure error growth across several chunks. For longer rollouts, supply the matching recorded states and explicitly disclose that assistance. Correct directional responses and lower error than the controls would support useful action conditioning; visually similar predictions despite different actions would undermine it. E03E16E18
8.3 Reading coverage
Visual audit: The title/author/version page, all Figures 1–10, all Tables 1–5, Equation (1), flow-matching equations (2)–(4), and the LAM objective were visually inspected. Each of the six final original-PDF crops was then inspected. Figure 10 is excerpted to its header and first three complete embodiment rows; the full ten-row page was also inspected. Table 1 retains its DreamZero* qualification and Table 2 retains trial-count context. All seven text chunks, covering the 27-page paper and references, were read. No appendix is present; separate supplements remain unverified. A subsequent independent repair pass also visually inspected PDF pages 13–14: Stage II/III supervision and Section 4.2, including the post-training action-loss weight 0.95 on page 14. These source pages are included in the next independent image review.
PDF pages inspected for this edition: 1, 2, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20. 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; arXiv:2608.27033v1, 27 August 2026
- 1. Introduction
- 2. Data Infrastructure
- 2.1. Multi-Source Embodied Experience
- 2.2. Unified Embodied Data Engine
- 2.3. Progressive Supervision and Data Balancing
- 3. Riemann-1.0 Design
- 3.1. Preliminary
- 3.2. Riemann-1.0: Fully Causal Action-Video World Model
- 4. Training Recipe
- 4.1. Pretraining
- 4.2. Post-training
- 5. Experiments
- 5.1. Real-World Experiments
- 5.1.1. Post-Training Data and Setup
- 5.1.2. Real-World Evaluation
- 5.1.3. Compositional Generalization and Out-of-Domain Evaluation
- 5.2. Simulation Evaluation
- 5.2.1. RoboCasa365
- 5.2.2. RoboTwin 2.0
- 5.2.3. LIBERO
- 5.3. Action-Conditioned Visual Rollout as a Multi-Embodiment Simulator
- 6. Related Work
- 7. Conclusion
- References
Outside the original text pass
- The inspected artifact is arXiv:2608.27033v1, 27 August 2026. Its title and all 16 authors match the catalog; Riemann Dynamics is verified on the title page although the catalog affiliation was not identified. No other revision or edition was supplied or compared.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction limitation was resolved by visually inspecting all ten figures, all five tables, and the principal equations in the supplied PDF. Previously concatenated table cells are legible in the original layout.
- Separate supplemental material availability has not been fully verified.
- No appendix or dedicated limitations section appears in the supplied 27-page PDF. References on pages 23–27 were read, but their cited works were not independently reviewed.
- No code or external project materials were inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
E01PDF p. 1, title, author block, and arXiv version line
The title page reads Riemann-1.0: An Embodied World Action Model for Physical AI; arXiv:2608.27033v1 [cs.RO], 27 Aug 2026. Its 16 named authors match the catalog in order. The displayed affiliation is Riemann Dynamics.
Go to primary source ↓E02PDF pp. 4–6, Sections 2.1–2.3
The corpus includes 200K+ hours of human video, 12K+ handheld/wearable hours, and 20K+ robot hours. Processing standardizes action-level trajectories using semantic annotation, hand/camera reconstruction, alignment, normalization, validity masks, filtering, and semantic-aware sampling.
Go to primary source ↓E03PDF pp. 7–8, Section 3.2, Equation (1) and adjacent paragraphs
The action conditional precedes the visual conditional, which additionally receives the current action. States are externally observed or replayed, not generated. Policy mode uses actual observations; simulator mode recursively uses predicted visual latents.
Go to primary source ↓E04PDF p. 8, Figure 4; pp. 8–9, Section 3.2, input conditioning and embodiment-specific interfaces
T5 encodes structured prompts; Wan VAE and 3D patch embedding encode packed views. Embodiment IDs select state/action projections and action heads around one shared Action/Video DiT backbone, with a shared latent velocity head.
Go to primary source ↓E05PDF p. 9, Figure 5 and token alignment; p. 10, Section 3.2, causal attention and loss masking
Clean history and noisy targets share a sequence. Action chunks align with latent transitions; attention excludes unavailable future clean tokens. Validity masks remove padded or invalid targets, and inactive padded channels remain zero during noise injection.
Go to primary source ↓E06PDF p. 7, flow-matching preliminary; p. 10, Equations (2)–(4)
The noise interpolation has target velocity ε−x. Separate latent and action velocity losses use their respective causal histories and combine as (1−λ)L_z+λL_a.
Go to primary source ↓E07PDF pp. 10–11, Section 3.2, Autoregressive online inference and Figure 6
The cache starts from text, initial observation, and state. A denoised action chunk executes for apf steps; actual observations and state update the cache. A full cache resets around the latest observation.
Go to primary source ↓E08PDF pp. 11–13, Sections 4 and 4.1, Stage I, Figure 7, and LAM objective
A separately trained, frozen latent-action VAE supplies posterior-mean pseudo actions from adjacent frames through a 32-dimensional bottleneck. Stage I groups these into chunks and uses λ=0.1; the text explicitly excludes deployable-policy learning as its purpose.
Go to primary source ↓E09PDF p. 12, Figure 7; p. 13, Section 4.1, Stage II and Stage III
Stage II combines UMI, robot, and annotated human-hand trajectories with λ=0.5. Stage III retains robot demonstrations only and raises λ to 0.9 while preserving visual supervision.
Go to primary source ↓E10PDF pp. 13–14, Section 4.2 and Section 5.1.1
Post-training combines four tasks into one generalist model with action weight 0.95. Section 4.2 gives three hours per task; Section 5.1.1 gives 15 demonstrations per task and identifies the Tianji Marvin dual-arm robot.
Go to primary source ↓E11PDF p. 14, baseline setup; p. 15, Table 1 (all task/average SR and PSR columns, caption) and Section 5.1.2
Table 1 reports Riemann task SR/PSR pairs 85.0/91.6, 90.0/98.4, 85.0/92.5, 80.0/95.2, averaging 85.00/94.43. G0.5 averages 70.00/80.18, ties folding SR 85.0 and desk SR 80.0; its kitchen SR is 35.0. LingBot-VLA kitchen PSR is 80.2. SR measures final completion; PSR measures milestone progress. DreamZero* is the authors’ 5B reproduction.
Go to primary source ↓E12PDF pp. 15 and 17–18, Section 5.1.3 and Table 2
Table 2 gives Riemann 8/10 and 5/10 compositional successes (65.0% average), then 10/10 and 7/10 OOD successes (85.0% average); overall 75.0%. π0.5 averages 45.0/50.0/47.5 and LingBot-VA 15.0/30.0/22.5, respectively. Color, receptacle, and order requirements change together. OOD exclusion concerns real-world post-training; no additional demonstrations or updates are used. Figure 9 says basket for the Rubik’s-cube example, while Table 2 and prose say storage box.
Go to primary source ↓E13PDF p. 18, Section 5.2 temporal interface, Section 5.2.1, and Table 3 (all category and Average columns)
Simulation uses 16 low-level steps per latent. RoboCasa uses 300-task pretraining then Target-50 fine-tuning. Riemann reports Atomic-Seen/Composite-Seen/Composite-Unseen 74.2/56.0/56.3 and average 62.6; ABot-M0.5 reports 70.6/44.3/45.6 and average 54.2. Composite gains are 11.7 and 10.7 percentage points.
Go to primary source ↓E14PDF p. 18, Section 5.2.2; p. 19, Table 4, Riemann and ABot-M0.5 rows, Clean/Randomized/Average columns
RoboTwin training covers 50 tasks with 50 clean plus 500 randomized demonstrations per task. Visually inspected Table 4 gives Riemann 94.6/94.0/94.3 and ABot-M0.5 94.0/94.2/94.1 for Clean/Randomized/Average, respectively.
Go to primary source ↓E15PDF p. 19, Table 5 (all suites and Average column), Section 5.2.3
LIBERO uses four suites with 10 tasks and 50 demonstrations per task. Riemann reports Spatial/Object/Goal/Long 99.6/100.0/97.6/98.6, average 99.0. Being-H0.5 and G0.5 average 98.9. Being-H0.5 Goal is 99.4 and G0.5 Long is 98.6, so Riemann does not strictly lead every suite.
Go to primary source ↓E16PDF pp. 20–21, Figure 10 and complete Section 5.3
Provided, policy-generated, or recorded actions condition visual-latent denoising; predicted or observed visuals can extend autoregressive history. Figure 10 illustrates heterogeneous embodiments. Section 5.3 supplies no quantitative visual-fidelity metric, changed-action control, or measured planning benefit.
Go to primary source ↓E17PDF pp. 8–13, implementation/training descriptions; pp. 14–21, complete experiment Sections 5.1–5.3; p. 22, conclusion
The method and training sections report interfaces and stage weights but omit Riemann parameter count, detailed optimizer/compute settings, LAM KL coefficient, denoising budget, and cache length. Experiments comprise task comparisons and qualitative rollouts, with no controlled mechanism ablation, inference timing, uncertainty estimates, or main four-task evaluation trial counts. No separate limitations section follows before References.
Go to primary source ↓E18PDF p. 20, Figure 10, header and first three rows: Agibot G1, BaiHu GR2, Galbot Yinhe
The figure labels uniformly sampled generated rollouts: Agibot G1 folding blue shorts, BaiHu GR2 picking up a carrot, and Galbot Yinhe folding and organizing clothes. The rows show successive generated camera frames; no ground-truth comparison or numeric action trajectory is displayed.
Go to primary source ↓8.5 Primary sources
Riemann-1.0: An Embodied World Action Model for Physical AI ↗
PDF · 10,953 extracted words
Source fingerprint
e6d34aaaeec29c8e41d0d8adfc08c35b4a582f0ba3b6a0313018e08244bab529