Causal World Modeling for Robot Control
1. Paper overview
In one sentence: LingBot-VA decodes robot actions from predicted visual futures inside a shared-attention model, trading costly video generation for partial denoising and feedback-grounded asynchronous execution. e03e04e08e09e14e15e19e20
| At a glance | What to know |
|---|---|
| Research problem | Author claim Manipulation under partial observability requires remembering earlier interactions while responding to execution errors. The authors argue that reactive observation-to-action learning entangles perception, dynamics and control; their alternative makes visual prediction an explicit inference-time input to inverse dynamics. e02e04 |
| Core mechanism | Reader analysis A video-action Mixture-of-Transformers interleaves modalities and shares attention at every layer while retaining separate modality parameters. This provides architectural integration beyond merely co-training independent models. e03 |
| A key reported result | RoboTwin 2.0: 50-task average: 92.93 Easy; 91.55 Hard. Reported average success rate (%). Multi-task training on 2,500 clean plus 25,000 randomized demonstrations; Easy uses fixed configurations and Hard randomizes poses/layouts. Table 1: Motus 88.7/87.0; π0.5 82.7/76.8. Horizon-3 success is 93.22/93.28 versus Motus 85.0/84.2. The reported advantage is larger in the longest horizon group, but this comparison does not isolate memory or causal masking. No uncertainty is given for these rows. e13e14 |
| Reading caution | Reader analysis Causality is across chunks: Equation (7) permits bidirectional attention within a chunk, and Figure 3 contains diagonal/group attention despite its preceding-token-only caption. This is a temporal generation constraint, not evidence that causal physical laws were identified. e05 |
Core contributions
- Reader analysis
A video-action Mixture-of-Transformers interleaves modalities and shares attention at every layer while retaining separate modality parameters. This provides architectural integration beyond merely co-training independent models. e03
Figure 2. Predicted visual transitions condition action decoding, with observed execution supplying the next context. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the lower-left task prompt and initial observation. Follow the video blocks to the imagined frames above them, then the action blocks to the corresponding predicted commands. The observations beneath later blocks represent the feedback that the next rollout can use. These repeated blocks illustrate successive stages; Section 3.3 describes separate modality parameters coupled by joint attention at every layer. The caption explicitly says video prediction precedes inverse-dynamics decoding, matching Algorithm 1. Thus the figure should be read as a unified video-action model with an ordered inference procedure, even though the introduction also uses the word simultaneous. e03e04e05e08e26
What it supports. The architectural connection is shared attention, not simply the fact that two objectives are trained together. Predicted future latents become inputs to action inference, so world modeling participates in control at deployment. Real observations subsequently replace imagined history in the synchronous loop, giving later predictions information about execution errors.
Where the evidence stops. This schematic does not specify the full attention mask or timing. Section 3.2 permits bidirectional attention within chunks; the introduction’s simultaneous-decoding wording differs from the caption and Algorithm 1. These details cannot be inferred from block placement alone.
2. Motivation
2.1 The problem and the proposed response
Manipulation under partial observability requires remembering earlier interactions while responding to execution errors. The authors argue that reactive observation-to-action learning entangles perception, dynamics and control; their alternative makes visual prediction an explicit inference-time input to inverse dynamics. e02e04
2.2 What this reading follows
A useful robot world model must help choose actions while remaining responsive to what actually happened. LingBot-VA couples a pretrained video stream to a smaller action stream through shared attention, then uses predicted visual transitions to infer motor commands. Its deployment design addresses a practical trap: computing ahead can make the model trust an obsolete imagined future. A forward-dynamics refresh reconnects that future to recent observations and the action currently executing. The reported simulation scores are strong, but understanding the paper also requires separating unified architecture from sequential inference, and checking its broad performance claims against the appendix’s less uniform results. e03e04e08e09e14e15e19e20
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 | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Conflicts with the recorded classification
WAM and IDM are supported: predicted future visual states directly condition actions. The recorded Dual-system architecture is less consistent with modality blocks fused through shared attention at every layer in one interleaved model. My assessment favors One Model × IDM; conceptual two-stage inference and asynchronous execution do not establish independent learned systems. The catalog snapshot is preserved. e03e04e08
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 From a visual future to an executable action
The conceptual decomposition starts with a history of observations and actions. The video model predicts the next visual chunk; inverse dynamics then asks which actions could realize that predicted transition from the current state. Equations (8) and (9) make the extra conditioning explicit: the action model sees the imagined future as well as both histories. LingBot-VA implements these roles with separate video and action transformer parameters that exchange information through joint attention at each layer. This lets the video backbone contribute learned visual dynamics while the narrower action stream specializes in motor outputs. Algorithm 1 still orders video generation before action generation. After execution, the loop encodes actual observations for later context. Understanding these separate levels—conceptual role, architectural coupling and inference order—prevents the two-stream design from being mistaken for two independent deployed models. e03e04e08e26
Figure 4. Forward dynamics refreshes the visual context while the robot continues its current action chunk. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two left panels as timelines: synchronous prediction creates the red hatched idle interval, while asynchronous prediction overlaps robot execution. The two right panels explain what changes inside that overlap. Naive async caches a real observation alongside an older visual forecast. In the grounded version, the FDM arrow takes recent feedback and the executing action into a new visual estimate, then inserts that estimate into the prediction context. The refreshed state is drawn without a hat, but the diagram labels it an updated prediction and Algorithm 2 computes it with FDM; it is not an observation from the future. e09e16e22
What it supports. The source’s mechanism is feedback-conditioned forecasting during execution. This addresses a specific failure mode: a video model may continue its own smooth imagined trajectory despite newer physical evidence. Table 3 reports higher success for the grounded variant than naive async, while still leaving a gap to the baseline.
Where the evidence stops. Figure indices are shifted by one step relative to the prose. Cache retention also remains ambiguous: Section 3.4 restricts recent history, but Algorithm 2 accumulates entries. The arrows establish information flow, not measured latency or guaranteed full-history memory.
5.2 Why incomplete denoising can still support control
Flow matching learns a velocity field that moves noisy samples toward training data. LingBot-VA applies this idea to both video and action tokens, but asks different things of their final representations. An executable action must be decoded completely; a visual latent only needs to contain enough information to condition that action. Noisy History Augmentation exposes the action model to partially corrupted visual context during training. The reader’s interpretation is that this can make action prediction less dependent on a fully reconstructed visual future, although the paper does not isolate the augmentation’s causal effect in a success-rate ablation. Algorithm 1 stops video flow at 0.5; the implementation instead lists three Euler steps ending at 0.6. Preserve that discrepancy when reproducing the claimed speed-quality tradeoff. Separately, causal ordering applies across chunks, with within-chunk attention explicitly allowed. e05e06e07e08e11e16
5.3 Make parallel prediction respond to physical feedback
Asynchronous control creates a mismatch between what the robot is doing and what the model can already observe. While one action chunk executes, prediction for the next chunk must begin before the current outcome is available. Simply retaining the old imagined outcome risks extending a hallucinated trajectory. Algorithm 2 therefore consumes queued observations from completed execution, adds the action now being executed to a temporary cache, and applies forward dynamics to estimate its outcome again. Next-chunk video and action prediction use this refreshed context. The FDM output remains a prediction grounded in recent feedback, not a new sensor measurement. Table 3’s grounded-versus-naive comparison is the relevant behavioral diagnostic, but it reports success only. Its remaining gap to baseline and the unclear cache-retention policy leave concrete questions about the cost and robustness of this pipeline. e09e16e22
5.4 Training and inference
During training
Teacher forcing supplies demonstration context. Video and action velocity losses are jointly optimized. With probability 0.5, video history is interpolated with Gaussian noise using an augmentation level uniformly sampled from 0.5 to 1. Action weights are initialized by interpolating video weights and scaling by the square root of the video/action width ratio. e06e07e17
The approximately 16K-hour corpus combines Agibot, RoboMind, InternData-A1, the OpenVLA OXE subset, UMI excluding DexUMI, RoboCOIN and internal demonstrations, with per-dataset 90/10 training/validation splits. Pretraining uses 1.4T tokens, AdamW, peak learning rate 0.0001, weight decay 0.01 and bfloat16; inverse-loss weight is 1. e10e12
Implementation samples chunk sizes from 1–4; the earlier 1–8 range is illustrative. General post-training uses 3K steps at 0.00001, but real-world evaluation specifies 500 steps at 0.0001 and sequence length 150,000. These schedules should remain distinct. e06e11e12e18
During inference
Algorithm 1 caches observed latents and executed actions, partially generates video, fully generates actions, executes them and encodes new observations. Section 4.2 specifies Euler integration: three video steps to flow time 0.6 and ten action steps to 1, with guidance scales 5 and 1. Algorithm 1 instead stops video integration at 0.5. e08e11
Asynchronous execution overlaps current motor commands with next-chunk prediction. Algorithm 2 incorporates queued feedback, conditions a forward-dynamics pass on the current action chunk, and uses its refreshed visual estimate to predict the next chunk. Post-training adds a forward-dynamics loss. The refreshed state remains imagined until execution feedback arrives. e09
5.5 Implementation flow
- Encode and align
A causal video VAE maps observations to latent tokens. Video is temporally sparsified by four; each retained frame aligns with four actions. The action vector contains end-effector position/quaternion, up to seven joints and a gripper value per arm; missing joint dimensions are padded. e04e03e10
- Fuse modalities
The video stream starts from Wan2.2-5B; the narrower action stream has the same depth. Separate QKV projections feed joint attention after dimensional alignment, followed by projection back to the action space. The implementation reports 30 layers, widths 3072/768 and approximately 5.3B total parameters. Frozen T5 embeddings provide instruction conditioning. e03e11
- Predict a transition and infer control
Video prediction uses latent observation and action history; inverse dynamics additionally consumes the predicted future. Despite introductory language about simultaneous decoding, Figure 2 and Algorithm 1 specify video generation followed by action generation. The commands are executed by the robot; imagined frames remain predictions. e03e04e08e26
6. Experiments & results
LingBot-VA turns a pretrained video generator into a robot policy: predict a short visual future, decode actions from that future and observation/action history, execute, and incorporate physical feedback. A unified Mixture-of-Transformers shares attention between video and action streams. Strong reported simulation results coexist with unresolved scoring and implementation inconsistencies.
6.1 Read the original evidence
Table 1. RoboTwin reports an advantage in both settings, with larger gains in the horizon-three group. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each method has paired Easy and Hard columns. Easy fixes initial configurations; Hard varies object poses and scene layouts. Read across a row before moving between horizons, since the rows group different tasks rather than extending a single episode. The last row is the reported 50-task average. Source note: the asterisk marks X-VLA results adopted from Motus; underlining marks the second-best entry and parentheses show gains over it. The main experiment trains on both clean and heavily randomized demonstrations. Its reported averages should not be substituted for Table 3’s separate deployment or initialization comparisons. e13e14e16
What it supports. LingBot-VA reports 92.93% Easy and 91.55% Hard, compared with Motus at 88.7% and 87.0% in this table. For horizon three, the entries are 93.22% and 93.28%, versus 85.0% and 84.2%. These support a reported advantage on more sequential tasks within this benchmark.
Where the evidence stops. The table provides no uncertainty or matched-compute control. Horizon-group improvements do not isolate the cache or attention mechanism. Imported baseline entries also limit claims that every model was rerun under an independently verified identical protocol.
Table 2. High average success coexists with a weaker Goal score and a tied Object result. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the four task-suite columns, then use the final average column for the overall comparison. The bottom row includes the paper’s reported variation terms; the table does not say whether these are standard deviations or another statistic. Section 4.3.2 describes three seeds and 500 evaluation trials per suite per seed. Dashes in earlier rows indicate missing entries and should not be treated as failures. Inspect Goal separately: OpenVLA-OFT reaches 97.9 while LingBot-VA reaches 97.2. Object is also not uniquely best, since CronusVLA has the same reported 99.6 score. Baseline values are adopted from reference [93]. e13e15
What it supports. The reported overall score is 98.5%, compared with X-VLA’s 98.1%. The Long suite reaches 98.5 ± 0.5%, versus X-VLA’s 97.6%. These are simulation policy evaluations and therefore stronger evidence for task execution than a visually plausible generated video alone, within the stated simulator protocol.
Where the evidence stops. Baseline training and evaluation were not independently verified here, and the ± terms have no defined statistic. Small differences near the success ceiling should not be presented as established statistical significance or uniform superiority across all suites.
Figure 5 (quantitative panels). Physical-robot gains vary by task and by the metric used to summarize execution. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the shared task labels to compare the top progress-score chart with the bottom success-rate chart. Yellow is LingBot-VA and green is π0.5; the figure states 50 real-world demonstrations per task. Appendix A describes 20 alternating trials per method. Progress averages intermediate-step credit, with half credit for a retry, whereas the stated success rule requires the maximum total score. This distinction matters when a policy completes much of a task but does not finish cleanly. The exact percentages discussed here come from Tables S2–S7, whose summary rows were visually checked, rather than estimated from the heights of unlabeled bars. e18e19e20
What it supports. Make Breakfast reports progress of 97.0% versus 73.0%, and Fold Pants 76.7% versus 30.0%. Fold Clothes reverses direction: 48.8% versus 62.9%, despite success labels of 35% versus 30%. Thus the plotted evidence itself contradicts the main text’s claim of improvement on every task and both metrics.
Where the evidence stops. Some appendix success flags violate the stated rule: Table S2 marks baseline trial 2 successful at 9/10 progress. Preserve the reported rates, but do not treat them as consistently defined completion counts. The full figure’s caption also swaps two task categories.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0: 50-task average Multi-task training on 2,500 clean plus 25,000 randomized demonstrations; Easy uses fixed configurations and Hard randomizes poses/layouts. | 92.93 Easy; 91.55 Hard. Reported average success rate (%) | Table 1: Motus 88.7/87.0; π0.5 82.7/76.8. Horizon-3 success is 93.22/93.28 versus Motus 85.0/84.2. The reported advantage is larger in the longest horizon group, but this comparison does not isolate memory or causal masking. No uncertainty is given for these rows. e13e14 |
| LIBERO: four-suite evaluation Four suites, ten tasks and 50 demonstrations per task before failure filtering; 4K fine-tuning steps; three seeds, 500 evaluation trials per seed per suite. | Spatial 98.5 ± 0.3; Object 99.6 ± 0.3; Goal 97.2 ± 0.2; Long 98.5 ± 0.5; overall 98.5. Reported success rate (%) | X-VLA: overall 98.1, Long 97.6. OpenVLA-OFT reaches 97.9 on Goal; CronusVLA also reaches 99.6 on Object. Overall and Long results lead the listed comparisons; Goal does not. Baselines are imported, and the meaning of the ± statistic is not defined. e13e15 |
| RoboTwin 2.0 Easy: asynchronous deployment ablation Table 3 deployment variants and WAN initialization comparison; no timing column or uncertainty. | Baseline 92.9; FDM-grounded async 90.4; naive async 74.3; WAN initialization 80.6. Reported success rate (%) | Horizon 3: baseline 93.2, grounded async 85.6, naive async 32.9, WAN 67.6. Grounding recovers much of naive async degradation while remaining below baseline. Prose claims twice-faster completion, but the table supplies no timing measurements. e16 |
| Six real-world manipulation tasks 50 demonstrations per task; 20 alternating trials per method. Progress normalizes graded intermediate steps; success rates below retain the appendix labels. | Breakfast 97.0/75; Delivery 84.5/65; Tubes 85.8/40; Screws 82.5/70; Clothes 48.8/35; Pants 76.7/70. Reported progress score / success rate (%) | π0.5: Breakfast 73.0/70; Delivery 73.0/25; Tubes 79.2/30; Screws 74.0/50; Clothes 62.9/30; Pants 30.0/30. Physical execution is demonstrated, with a Fold Clothes progress regression. Some success flags contradict the maximum-score rule, so these success rates are reported labels rather than consistently verified completion statistics. e18e19e20 |
| Post-training sample efficiency Figure 8, ten-demonstration regime; separate Make Breakfast and RoboTwin Easy experiments. | Breakfast 61.1; RoboTwin Easy 58.2. Reported progress score (%) | π0.5: 45.5 and 50.7, respectively. Reader subtraction gives gains of 15.6 and 7.5 percentage points. Prose assigns a 10.3-point RoboTwin gain to ten demonstrations, whereas that gap belongs to the five-demo bars (46.6 versus 36.3). e21 |
| Temporal memory: Wipe Plate and Search Box Exactly six wipes; sequential box search with the block always in the left box at test time. | 100 on both tasks. Reported success rate (%) | π0.5: 47 on Wipe Plate and 50 on Search Box. Useful behavioral diagnostics, but no isolated cache-removal comparison or task-specific trial counts establish the mechanism or uncertainty. e22 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Grounding substantially improves asynchronous success, but does not fully recover baseline performance. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two Deployment rows together: their difference concerns how asynchronous visual context is refreshed. Then compare both with the baseline row to see what is lost relative to the paper’s reference deployment. The Pretrain row instead changes initialization to WAN and belongs to a different comparison. Easy-all summarizes tasks; the rightmost column highlights horizon-three behavior. The source caption advertises world-model AR versus bidirectional ablation, but no such row is present. Likewise, the surrounding prose claims twice-faster completion, yet every visible column reports success rates rather than time. Keep those missing measurements separate from the entries that can actually be checked. e09e16
What it supports. Grounded async reaches 90.4% overall versus naive async at 74.3%, while the baseline is 92.9%. Horizon-three scores are 85.6%, 32.9% and 93.2%, respectively. The numerical pattern supports the value of refreshed context, with an especially large recovery in the longest group and a remaining performance cost.
Where the evidence stops. This table cannot verify the claimed twofold speedup or isolate AR versus bidirectional modeling. Nearby prose also cites 92.10 Easy/91.12 Hard without matching rows. No timing budget, variance or compute-matched grounding control accompanies the table.
7. Analysis & limitations
7.1 What the evidence leaves open
Causality is across chunks: Equation (7) permits bidirectional attention within a chunk, and Figure 3 contains diagonal/group attention despite its preceding-token-only caption. This is a temporal generation constraint, not evidence that causal physical laws were identified. e05
Memory scope is unresolved: Section 3.4 says history before t−1 is discarded, whereas Algorithm 2 accumulates cache entries and Section 4.5.2 credits full-history retention. Figure 4 also uses a one-step-shifted timeline relative to surrounding prose. e09e22
Table 3 lacks the advertised AR-versus-bidirectional row and timing data. Its Easy baseline is 92.9, while nearby prose cites 92.10 Easy/91.12 Hard without corresponding rows. Figure 5 misassigns task categories in its caption and overstates uniform improvement. e16e20
Figure 10 offers qualitative novel-object and spatial generalization examples without quantitative rates. The authors identify video-compression cost and absent tactile, force and audio inputs as future directions. e23e25
7.2 Questions for discussion
- Does feedback-grounded asynchronous inference retain its advantage under matched compute and controlled execution disturbances?
- How much of the long-horizon gain survives when cache length and robot pretraining data are controlled?
- Would a consistently applied real-world completion rule change success-rate rankings?
8. Reproducibility audit
8.1 Requirements and known gaps
A reproduction needs the pretrained video backbone, joint robot-data checkpoint, frozen text encoder, normalization statistics and embodiment-specific command mapping. The PDF omits GPU model/count, wall-clock training cost, software versions, normalization quantile thresholds and a complete internal-data recipe. e10e11e12e24
Resolve the 0.5/0.6 cutoff, cache retention and FDM implementation before matching deployment results. Pre-register the success rule: Table S2 marks baseline trial 2 successful despite progress 9/10, contradicting Appendix A. Recompute success flags and progress separately. e08e09e11e18e20
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 feedback grounding from extra computation
Reader-proposed check, not performed: evaluate synchronous, naive asynchronous and FDM-grounded asynchronous variants from the same checkpoint on the same RoboTwin Easy tasks and initial states, stratified by the paper’s horizon groups. Hold chunk size, action solver, video cutoff and cache length fixed. Include a compute-matched control whose extra forward pass receives stale feedback. Introduce the same timed object displacement during execution in every condition. Record success, recovery after displacement, end-to-end completion time, robot idle time and observation age. Grounding is supported if fresh-feedback FDM improves recovery beyond the stale-feedback control at comparable latency. If the gain disappears with matched compute or only survives without disturbance, the claimed feedback mechanism is weakened. e09e13e14e16
Check 2: Test the partial-denoising assumption directly
Reader-proposed check, not performed: post-train matched models with and without the stated noisy-history augmentation, keeping initialization, demonstrations, optimization and deployment schedule fixed. Evaluate each with video flow cutoffs 0.5, 0.6 and 1.0 while holding the action solver fixed. Compare equal video evaluation counts first, then separately compare equal latency budgets so cutoff and computation are not confounded. Log task success, progress, latency and visual-prediction error against observations that arrive after execution. The mechanism predicts that augmentation reduces the control penalty of stopping early, even when visual reconstruction remains imperfect. Failure to preserve action performance under partial denoising would challenge the explanation offered for the efficiency strategy. e06e07e08e11e17
8.3 Reading coverage
Visual audit: Visually inspected the title/revision page, all ten figures, Tables 1–3 and S1–S7, method equations, both algorithms, implementation details and Appendix A. All six final original crops were separately viewed. Table 1's asterisk and comparison notation are explained in its reading guide using the source caption. Figure 3 was checked for diagonal/group attention; Figure 4 was checked against its caption, prose and Algorithm 2, including the un-hatted predicted state and shifted timeline. All pages supporting retained method, numerical, evaluation and reproduction claims are included. Reference-only pages 19–22 were read in the text chunks but were not part of the visual pass. No external supplements or code were inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 23, 24, 25, 26, 27, 28, 29, 30, 31. 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, abstract and revision identifier (p. 1)
- 1 Introduction (pp. 1–3)
- 2 Preliminary: flow matching and conditional video generation (pp. 3–4)
- 3.1–3.4: formulation, architecture, training, synchronous and asynchronous inference (pp. 4–9)
- 4.1–4.5: data, implementation, results, ablations, sample efficiency, memory and generalization (pp. 10–17)
- 5 Related Work and 6 Conclusion (pp. 17–18)
- References (pp. 18–23)
- Appendix A and Tables S1–S7 (pp. 23–31)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The supplied artifact is arXiv:2601.21998v2 [cs.CV], dated 22 March 2026; its title and all 12 authors match the catalog. The catalog submission date is 29 January 2026. No earlier revision was supplied, so changes from v1 cannot be established.
- Text extraction does not reconstruct figure images; the retained PDF was separately inspected for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Code, checkpoints, external project pages and cited works were not inspected; no experiments were reproduced.
- The title page displays organizational logos but no explicit author-to-institution affiliations; affiliations are omitted.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and arXiv margin
The exact title is Causal World Modeling for Robot Control. The twelve authors match the catalog in order. The margin identifies arXiv:2601.21998v2 [cs.CV], 22 March 2026. Author roles are marked; no explicit affiliation list appears.
Go to primary source ↓e02PDF pp. 1 and 4, Section 1 and Section 3.1, Eq. (6)
The authors motivate explicit visual dynamics under partial observability and conceptually decompose prediction into future observations followed by inverse dynamics.
Go to primary source ↓e03PDF p. 4, Figure 2 and caption; p. 6, Section 3.3, Architecture / MoT Block
Figure 2 describes future-video prediction followed by action decoding. Separate video/action backbones have distinct QKV projections but exchange information through joint attention at every layer, with action dimensional projection and a residual return.
Go to primary source ↓e04PDF pp. 5–6, Section 3.2, Eqs. (7)–(9), state encoding; p. 6, Video Sparsification
The causal VAE encodes observations. Future latents condition on visual and action history; inverse dynamics additionally conditions on predicted future latents. Four actions align to each temporally retained video frame.
Go to primary source ↓e05PDF p. 5, Eq. (7) and following sentence; p. 7, Figure 3, caption and teacher-forcing text
The formulation permits bidirectional attention within chunks while remaining causal across chunks. The mask depicts diagonal entries and within-action-group attention; its caption broadly says tokens attend only to preceding tokens.
Go to primary source ↓e06PDF p. 7, Section 3.3, Variable Chunk Size / Noisy History Augmentation, Eq. (10)
Variable chunk sizes are illustrated with 1–8 and deployment K=4. Video-history augmentation uses noise interpolation with probability 0.5 and augmentation flow time in [0.5,1]; method text proposes inference to 0.5.
Go to primary source ↓e07PDF p. 8, Training Objective, Eqs. (11)–(12)
Video and action flow-velocity losses condition on noisy visual history; inverse dynamics includes the next visual observation. The total is dynamics loss plus λ times inverse loss.
Go to primary source ↓e08PDF p. 8, Algorithm 1, lines 1–13
Inference first integrates video flow to 0.5, then action flow to 1, executes actions, encodes observations and updates the cache with observed latents and executed actions.
Go to primary source ↓e09PDF pp. 8–9, Section 3.4, Figure 4, Algorithm 2 and Eq. (13)
Queued feedback and the executing action condition a forward-dynamics refresh before next-chunk prediction. Eq. (13) adds a post-training FDM objective. Prose restricts recent context while pseudocode appends cache entries; Figure 4 indexes real feedback as z_t where prose uses z_(t−1). Refreshed z_(t+1) lacks a hat in the figure but is labeled an updated prediction.
Go to primary source ↓e10PDF p. 10, Section 4.1, action interface and training-data composition
The two-arm interface has 30 dimensions. The corpus totals approximately 16K hours, includes six named dataset groups and internal demonstrations, and is split 90/10 per dataset; UMI excludes DexUMI.
Go to primary source ↓e11PDF p. 10, Section 4.2, Implementation Details
Wan2.2-5B supplies a 30-layer, width-3072 video stream; width-768 action stream adds about 350M parameters for a 5.3B model. Frozen T5 supplies text conditioning. Training uses K in [1,4]. Euler inference uses three video steps to 0.6 and ten action steps to 1; guidance scales are 5 and 1. Quantile thresholds are unspecified.
Go to primary source ↓e12PDF p. 11, Section 4.2, Pre-Training / Post-Training Details
Pretraining spans 1.4T tokens with AdamW, learning rate 0.0001, weight decay 0.01, bfloat16 and inverse-loss weight 1. General post-training is 3K steps at 0.00001, alternatively 1K steps at 0.0001.
Go to primary source ↓e13PDF p. 13, Section 4.3.2, Experimental Setup
RoboTwin uses 2,500 clean and 25,000 randomized demonstrations, 50K steps at 0.00001 and 12.5-Hz video/50-Hz actions. LIBERO uses four ten-task suites, filters failed demonstrations, trains for 4K steps and evaluates three seeds with 500 trials per suite per seed.
Go to primary source ↓e14PDF p. 13, Table 1, rows and caption; p. 25, Table S1, Average row
Table 1 reports LingBot-VA 92.93/91.55 Easy/Hard and Motus 88.7/87.0; horizon-3 entries are 93.22/93.28 versus 85.0/84.2. X-VLA values are adopted from Motus, as marked by an asterisk. Parentheses indicate gains over underlined second-best entries. Table S1 provides additional baseline decimals.
Go to primary source ↓e15PDF p. 14, Table 2, LingBot-VA, X-VLA, OpenVLA-OFT and CronusVLA rows and caption
LingBot-VA has Spatial 98.5±0.3, Object 99.6±0.3, Goal 97.2±0.2, Long 98.5±0.5 and average 98.5. X-VLA averages 98.1; OpenVLA-OFT has Goal 97.9; CronusVLA ties Object 99.6. Baselines come from reference [93]; the ± statistic is undefined.
Go to primary source ↓e16PDF p. 14, Section 4.4; p. 15, Table 3, caption and all rows
Easy-all entries are baseline 92.9, grounded async 90.4, naive async 74.3 and WAN 80.6; horizon-3 entries are 93.2, 85.6, 32.9 and 67.6. The caption advertises AR/bidirectional ablation without such a row. Prose claims twice-faster completion and 92.10/91.12 pretraining results, but the table contains no timing or Hard columns.
Go to primary source ↓e17PDF p. 7, Action Network Initialization; p. 15, Figure 7 and associated text
Action initialization interpolates pretrained video weights and scales by sqrt(d_v/d_a). Copy Init has lower plotted losses and gradient norms than Random Init and Share Weights. This is a training diagnostic, not an executed-task success comparison.
Go to primary source ↓e18PDF p. 13, Section 4.3.1; pp. 23–24, Appendix A, protocol and metrics
Real-world adaptation specifies 50 demonstrations per task, 500 steps at 0.0001 and sequence length 150,000. Appendix A specifies 20 alternating trials per method, step scores 0/0.5/1, normalized progress and success requiring maximum total score.
Go to primary source ↓e19PDF pp. 26–31, Tables S2–S7, Progress Score and Success Rate summary rows
LingBot-VA versus π0.5 progress/success: Breakfast 97/75 vs 73/70; Screws 82.5/70 vs 74/50; Clothes 48.8/35 vs 62.9/30; Delivery 84.5/65 vs 73/25; Tubes 85.8/40 vs 79.2/30; Pants 76.7/70 vs 30/30.
Go to primary source ↓e20PDF p. 11, Figure 5 and caption; p. 13, results; p. 23, scoring rule; p. 26, Table S2, Ours trial 8 and π0.5 trial 2; p. 28, Table S4
Figure 5 and Table S4 show worse Fold Clothes progress, contradicting uniform-improvement prose. Figure 5 caption swaps Pick Screws and Unpack Delivery categories relative to its image/body. Table S2 marks Ours trial 8 successful at 9.5/10 and π0.5 trial 2 successful at 9/10, contradicting Appendix A.
Go to primary source ↓e21PDF p. 15, Section 4.5.1; p. 16, Figure 8, five- and ten-demonstration labels
At ten demonstrations Breakfast bars are 61.1/45.5 and RoboTwin Easy bars 58.2/50.7. RoboTwin five-demo bars are 46.6/36.3; prose assigns their 10.3-point gap to ten demonstrations. Both vertical axes are Progress Score.
Go to primary source ↓e22PDF p. 16, Figure 9 and Section 4.5.2
The graphic reports Wipe Plate 100 versus 47 and Search Box 100 versus 50. Wiping requires exactly six repetitions; box search has a left-box target at test time. Text credits full-history KV caching, but supplies neither a cache-removal ablation nor task-specific trial counts. Caption left/right ordering is reversed.
Go to primary source ↓e23PDF p. 17, Figure 10 and Section 4.5.3
Training uses one pick-and-place object or localized positions; qualitative examples test novel objects or placements outside that region. No quantitative rates accompany them.
Go to primary source ↓e24PDF pp. 10–11, implementation/training; pp. 23–31, Appendix A
Architecture and optimization are specified, but GPU model/count, software versions and wall-clock compute are absent. The appendix contains evaluation protocols and tables rather than hardware details or an internal-data reconstruction recipe.
Go to primary source ↓e25PDF p. 18, Section 6, Future Work
Future directions include more efficient video compression and tactile, force and audio sensing for complex contact dynamics.
Go to primary source ↓e26PDF p. 2, Introduction after Figure 1; p. 4, Figure 2 caption; p. 8, Algorithm 1
The introduction describes visual denoising and action decoding as simultaneous; Figure 2 says video prediction occurs first, and Algorithm 1 orders video generation before action generation.
Go to primary source ↓8.5 Primary sources
Causal World Modeling for Robot Control ↗
PDF · 14,940 extracted words
Source fingerprint
57bd993b298765c005fc143e2550243cecb1e76c0682653634a85d34045b0442