When to Trust Imagination: Adaptive Action Execution for World Action Models
1. Paper overview
In one sentence: FFDC uses a WAM’s imagined visual future to decide when to interrupt action execution, trading additional verification and corrective inference for better task success. e-backbonee-verifiere-attentione-simulatione-reale-timing-discrepancy
| At a glance | What to know |
|---|---|
| Research problem | Source description A fixed action chunk commits the robot for the same duration regardless of how reliably its predicted future matches physical execution. Short chunks repeatedly invoke an expensive WAM during predictable motion; long chunks can continue through contact errors. The paper asks whether a WAM’s own predicted visual future can help decide when to replan. e-problem |
| Core mechanism | Source description Future–reality verification makes execution length depend on a learned assessment of the remaining plan, using predicted visuals at inference time. e-verifiere-attention |
| A key reported result | RoboTwin random setting: all 50 tasks: FFDC-WAM: 88.20%; 16.1 s; 1.69 calls. Success rate (%); average completion time (s); average WAM inference calls. Online rollouts; 100 executions per task. Random backgrounds, clutter, height, and lighting. Base-Motus trains/executes 16 actions; LC variants train 64 and execute the indicated prefix. Base-Motus: 85.66%; 24.4 s; 5.47 calls. LC-64: 87.26%; 13.6 s; 1.56 calls. The Base-Motus comparison gives +2.54 percentage points, 34.02% less time, and 69.10% fewer calls. LC-64 is faster with slightly lower success. The Base comparison changes backbone training horizon as well as execution. e-setupe-protocole-simulation |
| Reading caution | Source description The authors acknowledge incomplete coverage of real-world deviations by binary supervision, untested verifier scale/capability tradeoffs, and the need to study threshold sensitivity. e-limitations |
Core contributions
- Source description
Future–reality verification makes execution length depend on a learned assessment of the remaining plan, using predicted visuals at inference time. e-verifiere-attention
- Source description
FFDC combines a local temporal attention mask with cached WAM features; a separate binary training set mixes valid rollouts and corrupted or failed segments. e-attentione-training
- Reader analysis
The authors introduce Mixture-of-Horizon Training to improve long-horizon coverage, although the explicit recipe specifies start-time sampling and terminal padding rather than a distribution over horizons. e-training
Figure 2. One expensive imagined rollout supplies repeated lightweight verification. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin at the left: the action, video, and understanding experts provide the plan, predicted visual features, and semantic tokens L. The real observation enters FFDC separately. Snowflakes mark the video encoder and VLM; the paper does not give a complete freezing schedule. In the center, masked attention and feed-forward layers update the token sequence, and CLS feeds the score head. At right, the red time marker separates historical predictions from future predictions. Read matrix rows as querying tokens and columns as visible tokens. Its green blocks show modality visibility; Section 3.2 additionally restricts future action/visual interactions by temporal alignment and a local window. e-backbonee-verifiere-attention
What it supports. The architecture places a learned verifier between plan generation and continued action execution. Cached predictions remain available while the real observation changes, enabling repeated checks without rerunning Motus. The resulting confidence controls whether to retain the remaining plan; FFDC itself does not generate replacement actions.
Where the evidence stops. The figure’s cache annotation omits L, which Section 3.2 includes. The coarse mask does not depict within-block temporal restrictions. How deeper cached features are refreshed when real observations change is unspecified; use the text’s temporal rule without treating the drawing as a complete implementation.
2. Motivation
2.1 The problem and the proposed response
A fixed action chunk commits the robot for the same duration regardless of how reliably its predicted future matches physical execution. Short chunks repeatedly invoke an expensive WAM during predictable motion; long chunks can continue through contact errors. The paper asks whether a WAM’s own predicted visual future can help decide when to replan. e-problem
2.2 What this reading follows
A robot that predicts a long action sequence still needs to decide how much of it to execute. FFDC-WAM turns the accompanying visual prediction into a reference for that decision. Motus proposes actions and future visual tokens; a smaller verifier compares this cached plan with the latest observation and instruction, then either continues execution or requests a new plan. The key distinction is between generating a plausible future and checking whether the remaining actions still suit the physical state. The experiments support a useful success–efficiency tradeoff, but missing implementation details and inconsistent timing entries limit precise reproduction. e-backbonee-verifiere-attentione-simulatione-reale-timing-discrepancy
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 | Not applicable |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
“Other mechanisms,” “Not applicable,” and “Outside quadrants” fit the new contribution: an execution verifier attached to a joint video/action WAM. It neither extracts actions by inverse dynamics nor establishes a new One Model architecture. Efficient inference and long-horizon execution are directly supported; “Policy post-training & WM-RL” is only partly supported by supervised verifier training, with no RL objective shown. e-backbonee-verifiere-training
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 Separate the action proposal from the decision to keep executing it
The starting point is a WAM that already predicts both actions and latent video from an observation and instruction. FFDC does not derive actions from a generated goal image through inverse dynamics. It evaluates a remaining action proposal using additional evidence: semantic tokens, the latest real observation, and the predicted visual sequence around the check time. The “past” predictions represent what the WAM expected by now, while future predictions describe what should happen next. Temporal attention restricts interactions among the future action and visual tokens; CLS collects the information for a binary confidence score. At inference, a score at or above 0.5 permits continuation, while a lower score requests another WAM plan. This distinction connects the architecture to an actual execution decision rather than treating visual prediction as task success. e-backbonee-verifiere-attention
Figure 3. Execution length changes across tasks and within a task. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row from left to right; C1, C2, and later labels mark executed chunks, while “Inf.” marks a WAM call. In Move can pot, Base uses chunks of 16, 16, and 15 steps, whereas the displayed FFDC rollout completes a 43-step chunk after one inference. In Hanging mug, follow the FFDC row through chunks of 64, 32, and 6 steps. The green score permits continuation; the red score of 9.5159e−5 triggers replanning under the 0.5 threshold. The bottom row illustrates the consequence of continuing the long plan without that check. Ellipses indicate omitted portions of the displayed trajectories. e-rolloutse-verifiere-protocole-simulation
What it supports. The examples illustrate the intended allocation of computation: retain a plan during predictable motion and interrupt it near a difficult interaction. In the shown mug episode, FFDC succeeds after replanning, while the unverified rollout fails. These images make the control decision tangible but do not establish its average effectiveness.
Where the evidence stops. Selected trajectories do not measure failure prevalence, false interruptions, or confidence calibration. Base and FFDC also use different backbone training horizons. Use Table 1 for aggregate comparisons and avoid interpreting these images as a controlled intervention with identical trajectories.
5.2 Understand what the training labels can and cannot teach
Two training problems support the execution loop. Motus learns action and video prediction, while the verifier learns whether a segment is executable. The documented long-horizon sampling chooses a conditioning start anywhere in an episode and repeats the final action or frame beyond the endpoint. This exposes the backbone to late-stage starts; despite the Mixture-of-Horizon name, a probability distribution over different horizons is not specified. For verification, demonstrations and successful rollouts provide positives, while failed rollouts and four action-corruption operators provide negatives. Reader analysis: this supervision could teach useful incompatibilities between the plan and observation, but could also reward shortcuts that identify synthetic corruptions. The reported input ablations support the value of visual inputs; they do not distinguish those explanations. Held-out natural failures would be a particularly informative test. e-backbonee-traininge-ablatione-limitations
5.3 Compare the right baselines before attributing the gains
The headline simulation comparison changes both training horizon and execution strategy: Base-Motus trains on 16-action chunks, whereas the long-chunk family trains on 64. Consequently, the full improvement over Base cannot be attributed solely to the verifier. LC-64 supplies a closer fixed-execution reference: on the random all-task average it attains 87.26% success, compared with FFDC’s 88.20%, but needs less time and fewer WAM calls. Reader analysis: this supports buying some robustness with selective extra computation, rather than winning every efficiency metric. The physical comparison reinforces that interpretation: FFDC raises the two-task average from 45% to 80% while increasing time and calls. Finally, input ablations assess information sources, not the temporal mask itself. The most useful follow-up holds the backbone, data, and computation budget constant. e-protocole-simulatione-reale-ablation
5.4 Training and inference
During training
The backbone uses summed action and video rectified flow-matching losses. For an episode of length T, sample a start s uniformly, clamp action indices s+i and visual indices s+jr to T, and repeat terminal values beyond the episode. This includes late-stage conditioning states in training. e-backbonee-training
Verifier positives come from demonstrations and a small number of successful rollouts. Negatives come from failed rollouts and synthetic temporal swaps, gripper-sign flips, Gaussian noise in the second half, and suffix scaling. Binary cross-entropy trains executability labels. Counts, corruption strengths, train/validation separation, and optimization settings are not supplied. Figure 2 marks the video encoder and VLM with snowflakes; a complete freezing schedule is absent. e-traininge-verifier
During inference
After WAM inference, predicted action/visual and semantic tokens are cached. Each check encodes the latest real observation and evaluates FFDC without a full WAM call. Continue when confidence is at least 0.5; otherwise replan from the latest observation. Physical execution follows the selected actions; the verifier predicts whether to retain them, not replacement controls. e-verifiere-attention
The text calls verification high-frequency but does not specify its actual interval, numeric candidate horizon k, frequency ratio r, layer count N, or local-window size. Exact per-layer cache reuse when real observations change is also unresolved. e-verifiere-attention
5.5 Implementation flow
- Generate the candidate future
Motus jointly predicts an action chunk and latent visual tokens from the current observation and instruction. Figure 2 shows action, video, and understanding experts linked by joint attention. This paper adds verification to that backbone. e-backbonee-verifier
- Align imagined and observed states
At check time, assemble semantic tokens L, predicted historical visual tokens, the latest real observation O_t, predicted future visuals, the remaining action segment, and a learnable CLS token. The ratio r aligns action and video frequencies. Historical visual tokens here are previous predictions, not a stored history of real images. e-verifier
- Score with restricted attention
Future action/visual tokens attend to context and temporally preceding or aligned candidate tokens within a local window. An N-layer Transformer aggregates through CLS, then an MLP and sigmoid produce confidence. The temporal restriction does not establish a causal guarantee of executability. e-attention
6. Experiments & results
FFDC-WAM adds a learned execution verifier to Motus. It compares the current observation with cached predicted visual dynamics, planned actions, and instruction features, then continues or interrupts the rollout. RoboTwin results favor adaptive execution over short chunks, while physical experiments gain success at increased time and inference cost.
6.1 Read the original evidence
Table 1. Adaptive execution improves mean success while retaining much of the long-chunk efficiency. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each method has three columns: success rate SR, time T in seconds, and WAM calls. Higher SR is preferable; lower T and calls indicate greater efficiency under the reported timing protocol. Compare within a row before crossing between hard, easy, and all-task averages. Base-Motus trains and executes 16 actions; the LC variants train with 64 but execute different prefixes. FFDC adds adaptive verification. Read Rand.avg for the headline random-setting result, then Rand.hard to see where success improves most. The evaluation covers 50 tasks with 100 executions per task; the random setting adds backgrounds, clutter, height variation, and lighting changes. e-setupe-protocole-simulatione-timing-discrepancy
What it supports. On Rand.avg, FFDC reaches 88.20% success with 16.1 s and 1.69 calls versus Base’s 85.66%, 24.4 s, and 5.47 calls. This is a 2.54-percentage-point success gain with 69.10% fewer calls. LC-64 remains faster at 13.6 s and 1.56 calls, with 87.26% success.
Where the evidence stops. LC-16’s 51.4 s random-easy entry is also printed in Table 5, yet Table 1 gives 29.4 s random-hard and 22.0 s overall. These cannot be an ordinary weighted average. No corrected values or uncertainty intervals are provided; retain the discrepancy when comparing timing.
Table 2. Physical success improves while the cost of correction rises. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the banana and carrot rows separately before using the Average row. The physical platform is an Astribot S1 with 25 degrees of freedom, and the comparison here is LC-16 rather than Base-Motus. FFDC reports 80% success on each task; LC-16 reports 50% for banana and 40% for carrot. The time and call columns show that FFDC’s improvement is accompanied by extra computation and longer execution. Figure 4 on the preceding page illustrates example replanning events, but it does not provide the population of trials behind these averages. The source does not state the real-world trial count. e-setupe-reale-simulation
What it supports. The two-task average increases from 45% to 80%, an absolute gain of 35 percentage points. Average time increases from 25.6 s to 28.1 s and calls from 14 to 16. The physical experiments therefore support improved reliability at additional execution cost, rather than the simultaneous speedup reported against Base-Motus in simulation.
Where the evidence stops. The paper’s “35%” phrasing denotes the absolute difference between these averages. Without trial counts, uncertainty estimates, or broader physical tasks, this table cannot establish statistical significance or general reliability across real-world manipulation.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin random setting: all 50 tasks Online rollouts; 100 executions per task. Random backgrounds, clutter, height, and lighting. Base-Motus trains/executes 16 actions; LC variants train 64 and execute the indicated prefix. | FFDC-WAM: 88.20%; 16.1 s; 1.69 calls. Success rate (%); average completion time (s); average WAM inference calls | Base-Motus: 85.66%; 24.4 s; 5.47 calls. LC-64: 87.26%; 13.6 s; 1.56 calls. The Base-Motus comparison gives +2.54 percentage points, 34.02% less time, and 69.10% fewer calls. LC-64 is faster with slightly lower success. The Base comparison changes backbone training horizon as well as execution. e-setupe-protocole-simulation |
| RoboTwin hard subsets and clean aggregate The paper selects five hard tasks using Base-Motus performance; Table 1 separates random and clean settings. | FFDC: random hard 76.40; clean hard 76.00; clean all-task average 88.90. Success rate (%) | Base-Motus: 54.20, 57.80, 87.66. LC-64: 73.00, 74.60, 88.46, respectively. Gains concentrate on the hard subset; they are not universal task improvements. Clean HandBlock falls from 87% with Base-Motus to 60% with FFDC in Table 6. e-protocole-simulatione-appendix-easy |
| Astribot S1 banana/carrot pick-and-place Two physical tasks on a 25-DoF robot; real-world trial counts are not reported. | FFDC: banana 80%, carrot 80%; average 80%, 28.1 s, 16 calls. Success rate (%); average completion time (s); average WAM inference calls | LC-16: banana 50%, carrot 40%; average 45%, 25.6 s, 14 calls. The average improvement is 35 percentage points, accompanied by more time and calls. The paper’s “35%” wording is an absolute rate difference; no uncertainty intervals are supplied. e-setupe-real |
| FFDC input ablation on five hard RoboTwin tasks One input removed at a time. Section 4.3 does not name the environment setting; the full-model task rows match Table 4’s random rows. | Full model 76.4; without predicted visuals 71.6; real observations 72.4; actions 73.4; language 74.8. Average success rate (%) | Removing predicted visuals produces the largest mean drop: 4.8 percentage points. All four inputs contribute under this protocol. This does not isolate the causal mask, caching, or horizon-training strategy; no repeated-seed uncertainty is reported. e-ablatione-appendix-hard |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Predicted visuals contribute the largest average benefit among the tested inputs. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the Average row and compare the full FFDC-WAM columns at the right with each removed-input variant. “w/o Und” removes language conditioning, “w/o Pred” predicted visual tokens, “w/o Real” the current real observation, and “w/o Action” the predicted action segment. All variants report the same three outcome types, so a success change can be considered alongside time and calls. Then inspect individual task rows: improvements in an average need not hold on every task. For example, ScanObj has 88% success without language and 84% with the full verifier, despite the full model’s higher average. e-ablatione-appendix-hard
What it supports. Full FFDC reaches 76.4% mean success. Removing predicted visuals reduces this to 71.6%, a 4.8-percentage-point drop; removing real observations gives 72.4%. These comparisons support using both imagined and actual visual information. They also show that a beneficial average contribution is compatible with task-specific regressions.
Where the evidence stops. Section 4.3 leaves the clean/random setting unstated; the full-model rows match Table 4’s random results. Input removal does not isolate the causal mask, cache design, or horizon training. Missing repeated-seed uncertainty prevents a significance claim about the observed differences.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors acknowledge incomplete coverage of real-world deviations by binary supervision, untested verifier scale/capability tradeoffs, and the need to study threshold sensitivity. e-limitations
Table 1 prints LC-16 random-easy time as 51.4 s, repeated in Table 5, but random overall time as 22.0 s despite random-hard time of 29.4 s. These cannot form an ordinary weighted average. Figure 1(c) labels its x-axis inference calls while its caption calls it completion time. Neither discrepancy is silently corrected. e-timing-discrepancye-overview-discrepancy
Only two physical tasks are evaluated. Missing trial counts, confidence intervals, and a definition of time averaging over successes versus failures limit robustness and efficiency conclusions. e-setupe-reale-simulation
7.2 Questions for discussion
- Does temporally aligned future vision remain useful after matching verifier capacity and WAM-call budgets?
- How does a threshold calibrated on synthetic corruptions behave on previously unseen physical failure modes?
8. Reproducibility audit
8.1 Requirements and known gaps
The reported implementation uses PyTorch, four 80-GB A100s for backbone training, one A100 for verifier training, and one A100 for evaluation. Reproduction requires the Motus configuration, aligned video/action demonstrations, and executable RoboTwin and robot interfaces. PyTorch version, dataset sizes/splits, training duration, and detailed hyperparameters are absent. e-setupe-traininge-verifier
Proposed checks: hold the 64-action backbone fixed while testing aligned versus mismatched predicted visuals at matched inference budgets; then sweep the verification threshold against fixed-prefix controls and measure both success and end-to-end latency. These are follow-ups, not completed reproductions. e-protocole-ablatione-limitations
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does correct temporal alignment add value beyond extra verifier inputs?
Reader-proposed check: freeze one 64-action Motus backbone and use identical episode-separated training splits for full FFDC and a matched-capacity verifier without predicted visuals. Add a diagnostic condition that shifts the predicted visual sequence by one visual timestep while leaving actions and observations unchanged. Fix the check interval, and tune thresholds on validation data to match average WAM-call budgets. Evaluate paired initial states on the five hard RoboTwin tasks, separating natural rollout failures from synthetic corruptions. Measure task success, failure-detection recall, and unnecessary replans. The alignment explanation is weakened if shifted predictions preserve the full model’s advantage or if the advantage disappears on natural failures. e-traininge-attentione-protocole-ablatione-appendix-hard
Check 2: Measure the threshold tradeoff with complete execution costs
Reader-proposed check: keep the long-chunk backbone, verifier weights, observation rate, and seeds fixed. Sweep thresholds 0.1, 0.3, 0.5, 0.7, and 0.9, alongside LC-16/32/48/64 controls, under clean and random RoboTwin conditions. Record per-episode success, action duration, observation encoding, verifier latency, WAM latency, calls, and executed chunk lengths; report both all-episode and successful-episode timing with uncertainty. Use validation tasks to choose a threshold before testing. A useful adaptive policy should improve success at a comparable total time or call budget. If its apparent advantage vanishes when verification and failed episodes are counted, the efficiency interpretation is weakened. Resolve the printed LC-16 timing discrepancy from raw measurements rather than assuming a correction. e-verifiere-setupe-protocole-simulatione-timing-discrepancye-limitations
8.3 Reading coverage
Visual audit: All 14 supplied PDF pages were rendered and visually inspected, including the title/version/authors on p. 1, Figure 1 on p. 2, the architecture and equations on pp. 4–6, Figures 3–4 and evaluation protocol on pp. 7–8, Tables 1–3 on pp. 8–9, the conclusion on p. 10, references on pp. 10–11, and Appendix A with Tables 4–6 on pp. 12–14. Every final crop was separately viewed at its original exported resolution. Figure 2’s arrows, modality mask, snowflakes, and cache annotation were checked against Eqs. (4)–(7) and Section 3.2. The cache annotation omits semantic tokens mentioned in the prose, and its modality mask does not specify the within-block temporal rule. Figure 1(c)’s axis/caption mismatch and the repeated LC-16 timing inconsistency are disclosed. No supplied PDF page remains outside this pass; external supplements and code remain uninspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14. 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
- Abstract
- 1 Introduction
- 2 Related work
- 3 Method
- 3.1 Preliminary
- 3.2 Future forward dynamics causal attention
- 3.3 Training strategy and dataset construction
- 4 Experiments
- 4.1 Experimental setups
- 4.2 Main results
- 4.3 Ablation study
- 5 Conclusion
- References
- A Technical appendices and supplementary material
- A.1 Limitations
- A.2 Additional experimental results
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-only figure omission above was resolved by visually inspecting all 14 PDF pages, Figures 1–4, Tables 1–6, and all five final crops. All five supplied text chunks were read completely, including references and Appendix A.
- The title page identifies arXiv:2605.06222v2 [cs.RO], 9 May 2026, and labels the work a preprint. The exact title and all seven authors match the catalog. Earlier revisions were not supplied, so differences from v1 cannot be established. The title page additionally lists Astribot, Shenzhen, China, omitted from the catalog affiliation string.
- Code, external references, and separate supplements were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title block and arXiv margin
Exact title and seven authors verified; arXiv:2605.06222v2 [cs.RO], 9 May 2026, Preprint. Affiliations include Southern University of Science and Technology, The University of Hong Kong, and Astribot.
Go to primary source ↓e-problemPDF pp. 2–4, Sections 1–3.1
Fixed chunks trade unnecessary inference against prolonged unreliable execution; WAM-predicted future observations motivate adaptive verification.
Go to primary source ↓e-backbonePDF p. 4, Section 3.1, Eqs. (1)–(3)
Motus jointly predicts action and latent visual sequences from observation and instruction; training sums action and video flow-matching losses.
Go to primary source ↓e-verifierPDF p. 5, Eq. (4), Figure 2, Section 3.2 and Eq. (5)
Continue for e_t≥0.5, replan below. Verifier inputs include semantics, predicted past/future visual tokens, real observation, actions, and CLS. Figure marks video encoder and VLM with snowflakes. Text caches L as well as predicted actions/visuals; figure cache label omits L.
Go to primary source ↓e-attentionPDF pp. 5–6, Section 3.2, mask definition and Eqs. (6)–(7)
N-layer Transformer uses temporal visibility plus local windows; CLS aggregation, MLP and sigmoid score remaining-action validity. Predicted tokens are cached; latest real observation is encoded at checks.
Go to primary source ↓e-trainingPDF p. 6, Section 3.3, Eqs. (8)–(9)
Uniform conditioning starts and terminal repetition construct sequences. Positive demonstrations/successful rollouts and failed/corrupted negatives train binary cross-entropy. Four corruption operators are described without numeric settings or sample counts.
Go to primary source ↓e-setupPDF p. 6, Section 4.1
PyTorch and Motus; backbone trained on four 80-GB A100s, verifier on one A100, evaluations on one A100. RoboTwin has 50 tasks, 100 runs per task; random setting adds four perturbation types. Physical tests use 25-DoF Astribot S1 on two tasks.
Go to primary source ↓e-protocolPDF p. 7, Section 4.2, first two paragraphs
Base-Motus trains/tests horizon 16; LC-16/32/48/64 train horizon 64 and execute prefixes. Five named hard tasks are selected relative to Base-Motus success; aggregates separate hard/easy and clean/random.
Go to primary source ↓e-simulationPDF p. 8, Table 1, random/clean average and hard rows; PDF p. 7, Section 4.2
Random averages: Base 85.66%, 24.4 s, 5.47 calls; FFDC 88.20%, 16.1 s, 1.69; LC-64 87.26%, 13.6 s, 1.56. FFDC hard success is 76.40% random, 76.00% clean. Clean average FFDC 88.90%, Base 87.66%, LC-64 88.46%.
Go to primary source ↓e-rolloutsPDF p. 7, Figure 3 and caption; PDF p. 8, Section 4.2, first two paragraphs
Illustrated Move can pot rollout uses one FFDC WAM inference versus three Base calls. Hanging mug shows FFDC chunks of 64, 32, and 6 steps with a low-score replan; the unverified long rollout fails.
Go to primary source ↓e-realPDF p. 8, Real-world experiments and Figure 4; PDF p. 9, Table 2
Banana: LC-16 50%, 25.6 s, 14 calls; FFDC 80%, 26.7 s, 15. Carrot: 40%, 25.6 s, 14 versus 80%, 29.5 s, 17. Average: 45%, 25.6 s, 14 versus 80%, 28.1 s, 16; trial counts absent.
Go to primary source ↓e-ablationPDF p. 9, Section 4.3 and Table 3, Average and ScanObj rows
Average success: full 76.4%, no language 74.8%, no predicted visuals 71.6%, no real observation 72.4%, no actions 73.4%. Full average time 20.5 s and calls 2.34. ScanObj no-language 88% exceeds full 84%.
Go to primary source ↓e-limitationsPDF p. 10, conclusion continuation; PDF p. 12, Appendix A.1
Authors identify incomplete failure-mode coverage, unstudied parameter-scale tradeoff, and fixed 0.5 threshold sensitivity as limitations.
Go to primary source ↓e-appendix-hardPDF p. 12, Appendix A.2, Table 4, five random task rows
FFDC random hard success rates 93, 44, 88, 73, 84 match Table 3. The appendix reports both clean and random hard-task outcomes.
Go to primary source ↓e-appendix-easyPDF pp. 13–14, Tables 5–6, HandBlock and Average rows
Appendix contains random/clean easy-task breakdowns. Clean HandBlock success is Base 87%, FFDC 60%; clean easy averages are 90.98% and 90.33%.
Go to primary source ↓e-timing-discrepancyPDF p. 8, Table 1, LC-16 T(s) column; PDF p. 13, Table 5, LC-16 Average and later task rows
LC-16 random-easy 51.4 s and random-hard 29.4 s conflict with random-average 22.0 s. Table 5 repeats 51.4 s and contains many identical LC-16 success/time entries. Corrected values are not supplied.
Go to primary source ↓e-overview-discrepancyPDF p. 2, Figure 1(c), x-axis and caption
The x-axis reads WAM inference calls, whereas the caption describes task completion time. Quantitative conclusions in this report use the original tables instead of the scatter plot.
Go to primary source ↓8.5 Primary sources
When to Trust Imagination: Adaptive Action Execution for World Action Models ↗
PDF · 7,319 extracted words
Source fingerprint
4727ddd9f22de8d6428d073c6c1c9e2d06516b1bbc51734424f9d274ece60d16