FACT: Failure-Aware Causal Training for World-Action Models
1. Paper overview
In one sentence: FACT uses a shared action-first transformer to learn failed consequences without imitating failed actions, improving control while leaving consequence scoring an optional computation cost. e-probleme-architecturee-inferencee-real-resultse-future
| At a glance | What to know |
|---|---|
| Research problem | Source description A model trained only on successful demonstrations can imagine success even under a bad action. FACT asks how failed rollouts can teach accurate consequences without teaching the policy to imitate those actions. e-problem |
| Core mechanism | Source description A clean teacher-forced action slot and attention mask separate action generation from consequence prediction inside a shared transformer. e-architecture |
| A key reported result | Seen real-world bimanual manipulation: Failure-trained FACT: 89; with optional four-candidate scoring: 92. Average task success (%). Five tasks; 20 trials per task/method. Cube tasks use 200 success demonstrations and about 30 failures each; remaining tasks use 50 demonstrations. Success-only FACT: 82; Motus: 64; π₀.₅: 88. Failure training adds 7 percentage points under single-sample control; optional scoring adds 3 more. Baselines share success fine-tuning data, but pretraining and failure supervision differ. e-real-protocole-real-resultse-hardware |
| Reading caution | Source description Broader robot/human data and improved progress estimators remain future work. Best-of-N adds a scoring pass; action-only control avoids that cost. e-limits |
Core contributions
- Source description
A clean teacher-forced action slot and attention mask separate action generation from consequence prediction inside a shared transformer. e-architecture
- Author claim
Failure-aware video and progress supervision improves policy success and prediction of failed outcomes in the reported settings. e-real-resultse-sime-future
Figure 2. One backbone learns behavior from successes and consequences from both outcomes. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the token row in panel (a). P supplies the observation prefix, A is the noisy action being learned, and G is a clean recorded action used to condition the world branch. The red 'no attn' marker separates noisy-action generation from that clean conditioning path. V and I represent value and future video. The enlarged block at right adds an action adapter after the feed-forward network for robot tokens. Now follow panel (b): successful demonstrations activate all three losses; failed rollouts cross out only the action loss. At deployment, the stage labels indicate that a generated clean action becomes the condition for optional consequence prediction. e-architecturee-targetse-traininge-inference
What it supports. The important shared component is the transformer itself. Failure video and value losses can affect parameters also used to generate actions, even though failed actions are excluded as imitation targets. This provides an architectural route for failure supervision to improve action-only control.
Where the evidence stops. The red marker denotes blocked attention, not a frozen backbone or stop-gradient. The failure-progress bar is schematic; Equation (6) defines the clipped target and the implementation sets the penalty to 1.
2. Motivation
2.1 The problem and the proposed response
A model trained only on successful demonstrations can imagine success even under a bad action. FACT asks how failed rollouts can teach accurate consequences without teaching the policy to imitate those actions. e-problem
2.2 What this reading follows
A missed grasp is poor imitation data but useful evidence about what an action causes. FACT builds that distinction into a video transformer: the policy learns actions from successful demonstrations, while the world branch learns the observed consequences of both successful and failed actions. A clean action condition connects each prediction to the behavior that produced it. This reading follows the attention mask, the policy comparisons and the failed-future diagnostics. The central distinction is between improving the policy through consequence supervision during training and spending extra computation to score alternative actions during deployment. The reported gains involve both, under separate evaluation settings. e-probleme-architecturee-inferencee-real-resultse-future
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
Shared transformer blocks and robot-token adapters support One Model. The action-first, action-conditioned world branch supports Other mechanisms / Outside quadrants rather than inverse dynamics or simultaneous action/future denoising at deployment. Joint video-action training and policy continuation are present; reinforcement-learning updates are only a future direction, not a demonstrated component of the broad post-training/WM-RL label. e-architecturee-inferencee-algorithmse-limits
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 Give a failed action two different roles
Consider a rollout in which the gripper misses a cube. The recorded action is still a valid explanation of the observed future, even though it is an undesirable action to imitate. FACT therefore puts that clean action into G for the world branch, while retaining a separate noisy action segment A for policy training. On a successful window, action, video and value losses all contribute. On a failed window, the action loss is masked and the other two remain. This distinction applies across failed-episode windows in Algorithm 1; the progress penalty separately depends on failure onset through Equation (6). Thus exclusion from imitation and the timing of the lowered value target are related design choices with different roles, not one undifferentiated failure label. e-architecturee-targetse-traininge-algorithms
Figure 3. The mask makes the action condition available to the world branch without revealing it to the policy. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as a query and each column as an accessible key. In training, the A row can use P and A but has an empty G cell, preventing access to the clean action target. The V and I rows have filled G cells and empty A cells, so consequence learning uses the recorded action rather than a noisy prediction. I can also attend to V. On the right, the dashed P/A block marks Stage 1. The column remains labeled A below that block, but the note explicitly says the denoised action then goes to G; Section 3.3 and Algorithm 3 identify it as the clean condition for Stage 2. e-architecturee-inferencee-algorithms
What it supports. The matrix explains why failed actions can remain informative training inputs without becoming action targets. At inference, FACT can finish after the dashed action block. World prediction is a later, optional computation using the clean generated action, rather than a prerequisite for producing that action.
Where the evidence stops. The inference panel compresses two stages and omits a separate G row/column. Its A label must not be read as world tokens attending to still-noisy actions; the adjacent note and Algorithm 3 specify the clean transfer.
5.2 Trace how consequence learning can change action-only behavior
The attention mask protects the action prediction from access to the clean action target. It does not create separate policy and world networks: both use the same transformer, with additional action capacity supplied by robot-token adapters. Consequently, as a reader interpretation of the architecture, video and value gradients can change representations used by later action generation even when no failed action is an imitation target. The empirical comparisons support this route without identifying exactly which representation changes. Success-only FACT falls from 82% to 58% on seen tasks when video co-training is removed; failure training raises the full model to 89%. The causal-mask ablation also removes G and changes joint denoising, so its 77% result cannot attribute the entire difference to one attention edge. e-architecturee-traininge-ablatione-real-results
5.3 Treat the progress head as a conditional ranking device
After Stage 1, a sampled action can be executed directly. Scoring instead inserts each clean candidate into the world-conditioning slot and predicts its progress value before choosing the highest-scoring action; rendering future video is optional. The target is normalized episode progress with a failure penalty, not a learned discounted return or an established probability of success. With the reported penalty of 1, Equation (6) assigns zero to endpoints marked failed. That creates a discriminative training signal, but ranking quality must still be tested. The real-world scoring gains and placement-grid diagnostic support usefulness in the evaluated cases. They do not establish calibration: Appendix H even reports unclipped negative predictions. Figure 7 also shows that increasing candidates beyond four need not improve the plotted completion rate. e-targetse-inferencee-real-resultse-valuee-scoring
5.4 Training and inference
During training
Initialize from WAN2.2-5B and use AdamW: learning rates are 2×10⁻⁴ for the action FFN and 2×10⁻⁵ for the backbone. Action/value/video loss weights are 20/1/1; H=48 and the failure penalty is 1. Video targets use offsets [0,H/4,H/2,3H/4,H]. e-training
Flow matching interpolates clean modality tokens with Gaussian noise and learns the noise-minus-clean velocity by squared error. Algorithm 2 first trains on successes, collects failed policy rollouts, then continues mixed training. Failed windows retain video/value losses and lose action imitation. e-traininge-algorithms
During inference
Stage 1 denoises actions from observation and state, normally for 20 flow-Euler steps, and can return immediately. Stage 2 conditions on the clean sampled action to predict value and optionally video. Candidate scoring ranks predicted consequences before executing the selected action; it does not require physically trying every candidate. e-inferencee-algorithms
Optional real-world evaluation samples four candidates. The same model scores them; no separate critic is trained. The source reports prefix key-value caching for action-only inference. e-inferencee-real-protocol
5.5 Implementation flow
- Encode the robot context
The observation combines three camera views and proprioception. Camera streams are packed into one video canvas, with per-view VAE latents precomputed; task prompts provide T5 embeddings for text cross-attention. e-formulatione-hardwaree-prompts
- Separate noisy actions from clean action conditions
Tokens are ordered P, A, G, V, I: observation prefix, noisy predicted action, clean recorded action, value and future video. A cannot attend to G; V and I attend to G rather than noisy A. This prevents the policy from seeing its clean target. e-architecture
- Share consequence learning with the policy
One causal video diffusion transformer processes the modalities, with an action adapter after each block's feed-forward network on robot tokens. World losses can update shared parameters used for action generation; masking action loss does not freeze the policy backbone. e-architecturee-training
6. Experiments & results
FACT learns actions and their consequences in one video transformer. Successful demonstrations supervise all outputs; failed rollouts supervise future video and task progress while their action-imitation loss is disabled. Action-only deployment skips future prediction, with optional action-conditioned scoring offering additional gains at extra computation.
6.1 Read the original evidence
Table 1. Video co-training and failure supervision each improve the reported RoboTwin average. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read downward within the FACT rows before comparing external systems. The no-video ablation at the bottom averages 81.8%; adding video co-training gives 85.6%; adding failed rollouts reaches 87.5%. The Clean and Rand. columns show that both additions improve the two aggregate settings. Section 4.1 reports about 1,300 failed rollouts. Appendix D specifies 50 tasks with 2,500 clean and 25,000 randomized training demonstrations, and Table 6 states 100 evaluation trials per task in each split. These appendix details define what the compact table averages; the external-baseline block contains differently pretrained systems, not additional FACT ablations. e-sime-sim-protocol
What it supports. The reported average rises by 3.8 percentage points with video co-training and another 1.9 with failure supervision. Failure-trained FACT approaches Motus, which remains slightly higher at 87.8%. These results concern executed simulated tasks; they are separate from future-image prediction quality.
Where the evidence stops. The table reports aggregates without uncertainty. Adding failure data changes the training information available, so this comparison alone does not isolate data quantity from consequence supervision. Per-task results include regressions as well as gains.
Tables 2–3. Failure training improves both reported averages; optional scoring adds a separate deployment gain. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Keep the left and right averages separate: Table 2 covers five seen tasks, while Table 3 covers three held-out object and instruction variants. Each task/method cell represents 20 trials. Within the FACT block, first compare Ours with Ours w/ failure to examine training, then compare the failure-trained row with its optional-scoring row to examine deployment. The optional rows use four candidates. Below the seen-task block, the ablations remove video, remove clean-action causal conditioning, or retain failed-action imitation. All external baselines were fine-tuned on the same expert success demonstrations, but they do not share identical pretraining or FACT's additional failure supervision. e-real-protocole-real-resultse-ablatione-hardware
What it supports. Seen-task averages rise from 82% to 89% with failure training and to 92% with scoring; unseen averages rise from 67% to 77% and 82%. The 63% failed-action-loss ablation supports separating consequence supervision from imitation. π₀.₅ still leads the unseen comparison at 85%.
Where the evidence stops. Section 4.2 specifies four candidates for optional rows but calls other variants single-sample, leaving the separate 'Ours + scoring' ablation's candidate count ambiguous. No uncertainty is reported, and shared success demonstrations do not equalize pretraining or failure data.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Seen real-world bimanual manipulation Five tasks; 20 trials per task/method. Cube tasks use 200 success demonstrations and about 30 failures each; remaining tasks use 50 demonstrations. | Failure-trained FACT: 89; with optional four-candidate scoring: 92. Average task success (%) | Success-only FACT: 82; Motus: 64; π₀.₅: 88. Failure training adds 7 percentage points under single-sample control; optional scoring adds 3 more. Baselines share success fine-tuning data, but pretraining and failure supervision differ. e-real-protocole-real-resultse-hardware |
| Unseen real-world object and instruction variants Held-out Stack Cubes, Pick Cubes and Stack Bowls variants change colors, shapes and instructions; 20 trials per cell. | Failure-trained FACT: 77; optional scoring: 82. Average task success (%) | Success-only FACT: 67; Motus: 62; π₀.₅: 85. Reported rounded averages improve, but FACT remains below π₀.₅ on these variants. e-real-protocole-real-results |
| RoboTwin clean and randomized manipulation 50 tasks; 2,500 clean and 25,000 randomized training demonstrations plus about 1,300 failures; 100 evaluation trials per task per split. | 88.4 / 86.6 / 87.5. Clean / randomized / average success (%) | Success-only FACT: 86.3 / 84.9 / 85.6; without video: 82.5 / 81.0 / 81.8; Motus: 88.7 / 87.0 / 87.8. Video and failure co-training each improve the average, while Motus retains a small aggregate lead. e-sime-sim-protocol |
| Seen-task mechanism ablations Same five-task real-world benchmark. The causal-mask ablation uses success-only data and removes G while jointly denoising outputs. | Without causal mask: 77; with failed-action imitation: 63; without video: 58. Average success (%) | Success-only FACT: 82; failure-trained FACT with masked action loss: 89; success-only scoring: 79. Results support clean action conditioning and selective failure supervision. The mask ablation changes conditioning as well as attention; it is not an isolated one-edge intervention. e-ablatione-real-protocol |
| Held-out future-image prediction 512 windows, evenly split between successful demonstrations and failed rollouts; both checkpoints use 20 denoising steps. | Failure subset: 25.92 / 0.8290; success subset: 26.08 / 0.8286; all: 26.00 / 0.8288. PSNR / SSIM (higher is better) | Success-only checkpoint: failure 19.51 / 0.7461; success 26.12 / 0.8285; all 22.82 / 0.7873. The gain concentrates on failed futures. Image similarity supports consequence prediction, not a robot-success or calibration guarantee. e-futuree-failed-image |
| Action-only deployment latency Measured on an RTX PRO 6000; failure-trained FACT uses its action-only path. | 380. Inference latency (ms) | Motus: 1220; Cosmos: 620; π₀.₅: 47; π₀: 45. FACT is faster than these WAM comparators but slower than the listed action policies. This measurement excludes optional candidate scoring. e-latency |
| Failure-data scaling on three RoboTwin clean tasks Handover Block, Open Microwave and Put Bottles Dustbin; use 0%, 50% or 100% of available failures. At 100%, failures comprise about 45% of training data. | 32.7 to 57.3 from zero to all available failures. Average success (%) | Figure 6 reports a monotonic average across the three fractions. This restricted experiment supports additional failure supervision; individual task bars are not uniformly monotonic. e-scaling |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. A bad action should predict its failed consequence even when the instruction requests success. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the current camera views at left to locate the red cube and plate. The next group is the observed future after the bad action, which serves as the comparison target. The success-only prediction is the following group; the paper's white dotted box highlights a grasp that did not match the recorded failed outcome. The rightmost prediction comes from failure-aware co-training and more closely follows the observed result. The caption holds the bad-action condition fixed across the two models. These panels illustrate the consequence-prediction branch, not a comparison of newly executed actions selected by the two policies. e-failed-imagee-future
What it supports. The example makes the learning target concrete: a failed action is paired with the future it actually produced. Failure-aware supervision changes the predicted consequence under that same action, supporting the paper's interpretation of reduced success-biased hallucination. The held-out metrics provide a broader check than this selected image.
Where the evidence stops. This is one qualitative example, not a measured hallucination frequency. Resemblance to an observed failed frame does not establish accurate dynamics for arbitrary actions or guarantee successful robot control.
Table 8. The improvement concentrates on failed futures while successful-future metrics remain nearly unchanged. Original paper, p. 19 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read this table by outcome subset before reading the pooled All row. Within each metric, Ours is the success-only checkpoint and Ours w/ failure adds failed-rollout supervision. The arrows indicate that larger values are better. Appendix F and the table caption specify 512 held-out future-prediction windows, divided equally between successful demonstrations and failures, with 20 denoising steps for both models. The highlighted failure-row baseline cells make the main weakness visible. Compare the successful-rollout row as a control: its similarity metrics barely change, so the pooled improvement should not be described as an equal gain across all kinds of futures. e-future
What it supports. On failed futures, PSNR rises from 19.51 to 25.92 and SSIM from 0.7461 to 0.8290. Successful-future PSNR changes from 26.12 to 26.08 and SSIM from 0.8285 to 0.8286. This supports a targeted improvement in failure consequence prediction rather than a broad numerical improvement on every subset.
Where the evidence stops. PSNR and SSIM measure image similarity, not value calibration or executed task success. The paper does not give confidence intervals or a detailed episode-level holdout construction, so independence of sampled windows cannot be established from this table.
7. Analysis & limitations
7.1 What the evidence leaves open
Broader robot/human data and improved progress estimators remain future work. Best-of-N adds a scoring pass; action-only control avoids that cost. e-limits
Only 20 real-world trials per cell are reported, without confidence intervals. Section 4.2 assigns N=4 to optional rows and calls other variants single-sample, leaving the separate 'Ours + scoring' ablation's candidate count ambiguous. e-real-protocole-ablation
Progress targets use episode time and annotated failure onset rather than calibrated success probabilities. Appendix H explicitly leaves diagnostic predictions unclipped, explaining negative heat-map scores despite bounded targets. Figure 7 falls from four to eight candidates, although adjacent prose describes a smaller gain. e-targetse-valuee-scoring
7.2 Questions for discussion
- How sensitive is ranking to episode-length normalization and uncertain failure-onset labels?
- Does failure supervision improve shared action representations when training compute and candidate pools are held fixed?
8. Reproducibility audit
8.1 Requirements and known gaps
Recreate the two YAM arms, D435 main camera, two D405 wrist cameras, GELLO collection, multi-view preprocessing and exact prompts. The PDF specifies learning rates and losses but omits batch size, training duration, training accelerator count and software versions. e-hardwaree-promptse-traininge-algorithms
Failure-onset annotation needs resolution: Section 3.3 assumes annotations, while Algorithm 2 says 'when available' without a fallback. Proposed checks should control data, compute and candidate pools when testing action-loss masking and value-based selection. e-targetse-algorithmse-ablatione-scoring
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: Hold failed rollouts fixed and test which loss may imitate them
Reader-proposed check, not performed: start from the same success-trained checkpoint and reuse identical successful and failed windows. Compare FACT's masked failed-action loss with an otherwise identical run that retains that loss; include a success-only continuation matched for optimizer steps and successful-window exposure. Keep the attention mask, loss weights, future offsets and evaluation seeds fixed. Evaluate action-only task success and held-out failure-future PSNR/SSIM, reporting uncertainty across seeds and trials. The selective-supervision hypothesis predicts better task success with masking while retaining useful failure-future prediction. If masking provides no repeatable advantage, the claimed separation between learning consequences and imitating failures would need qualification. e-traininge-algorithmse-ablatione-future
Check 2: Separate candidate availability from the value head's ability to rank
Reader-proposed check, not performed: cache identical candidate action pools for held-out RoboTwin states and score each pool with success-only and failure-trained checkpoints. Reset the simulator to each state before executing candidates to measure their outcomes. Compare first-candidate, random-candidate and value-max selection at one, four and eight candidates; include an outcome-based oracle only as an evaluation ceiling. Measure selection success, ranking errors and full scoring latency on one fixed accelerator. Also stratify errors by time since annotated failure onset. Better ranking from the failure-trained scorer on identical pools would support consequence discrimination; gains explained only by pool size, or systematic preference for bad candidates, would weaken that explanation. e-inferencee-targetse-scoringe-valuee-latency
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page; all method equations and masks; Figures 1–13; Tables 1–8; and Appendix A's algorithms. Appendix pages supplying hardware, prompts, dataset sizes, evaluation trials, latency and diagnostic details are included. All six final original crops were inspected; the attention crop was widened to retain both query labels. Figure 3's compressed inference notation was checked against Section 3.3 and Algorithm 3, and Figure 7's plotted decline beyond four candidates is preserved as a prose/plot discrepancy. References on pages 10–13 were read in the complete text but were not visually inspected. No separate supplement, code or linked resource was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 14, 15, 16, 17, 18, 19. 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 contribution note (p. 1): arXiv:2608.10232v1, 10 August 2026
- 1 Introduction (pp. 1–2)
- 2 Related Work (pp. 2–3)
- 3 Method, including 3.1–3.3 and Equations (1)–(10) (pp. 3–5)
- 4 Experiments, including 4.1–4.3 (pp. 6–8)
- 5 Limitations and 6 Conclusion (p. 9)
- References (pp. 10–13)
- Appendix A: Training and Inference Algorithms (p. 14)
- Appendix B: Task Overview (p. 15)
- Appendix C: Real-World Setup and Task Prompts (pp. 15–16)
- Appendices D–H: simulation details, latency, prediction metrics and value diagnostics (pp. 17–19)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The inspected title, five authors and affiliation match the catalog. This report concerns the supplied v1 dated 10 August 2026; no different revision or edition was supplied or compared.
- Text extraction does not reconstruct figure images; the retained PDF was therefore visually inspected for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Code, linked websites and external references 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, author footnote and arXiv margin
Title and all five authors match; UC San Diego affiliation; Peng and Liang contributed equally; v1 is dated 10 August 2026.
Go to primary source ↓e-problemPDF pp. 1–2, abstract and Section 1
Motivates learning failed consequences without imitating failed actions; introduces optional scoring.
Go to primary source ↓e-formulationPDF p. 3, Sections 3.1–3.2, Equations (1)–(4)
Defines instruction, multiview observation, proprioception, action horizon, future observations and normalized progress.
Go to primary source ↓e-architecturePDF p. 4, Figures 2–3, Equation (5), Section 3.2
Shared transformer with robot-token adapters; training V/I attend G, not A; A cannot see G. Inference diagram labels A but annotates transfer of the denoised action to G.
Go to primary source ↓e-targetsPDF p. 5, Section 3.3, Equation (6) and implementation details
Progress uses t/T, failure-onset indicator and clipped penalty; experimental failure penalty equals 1.
Go to primary source ↓e-trainingPDF p. 5, Section 3.3, Equations (7)–(9) and implementation details
Flow matching, success/failure losses, WAN2.2-5B initialization, AdamW, learning rates, loss weights, 48-step chunks and future offsets are specified.
Go to primary source ↓e-inferencePDF p. 5, two-stage inference, optional scoring and Equation (10)
Action denoising can run alone; optional world pass conditions on each clean candidate and selects maximal progress; prefix caching and 20 default steps are reported.
Go to primary source ↓e-algorithmsPDF p. 14, Appendix A, Algorithms 1–3
Success training precedes failed-rollout collection and co-training; failure actions are masked; onset is annotated when available; candidate scoring precedes selected action return.
Go to primary source ↓e-real-protocolPDF p. 6, Section 4.2; p. 15, Appendix B
Five seen and three unseen variants; demonstration/failure counts; 20 trials per cell; optional rows use four candidates while the prose calls other variants single-sample.
Go to primary source ↓e-real-resultsPDF p. 7, Tables 2–3, FACT and external-baseline rows, Avg. columns
Seen averages are 82/89/92 for FACT variants and 64/88 for Motus/π₀.₅; unseen averages are 67/77/82 and 62/85.
Go to primary source ↓e-simPDF p. 6, Section 4.1 and Table 1, Clean/Rand./Average columns
About 1.3K failures are added; FACT averages are 81.8 without video, 85.6 success-only and 87.5 with failures; Motus is 87.8.
Go to primary source ↓e-sim-protocolPDF p. 17, Appendix D; p. 18, Table 6 and caption
Training totals are 2,500 clean and 25,000 randomized demonstrations across 50 tasks; evaluation uses 100 trials per task/split. Per-task results include both gains and regressions.
Go to primary source ↓e-ablationPDF p. 7, Table 2 ablation rows and Section 4.3 policy-performance paragraph
Seen averages: scoring alone 79, no causal mask 77, failed-action loss 63, no video 58. The causal-mask variant removes clean G and jointly denoises outputs without failure data.
Go to primary source ↓e-futurePDF p. 7, Table 4; p. 17, Appendix F; p. 19, Table 8, all rows and metric columns
512 held-out windows split equally by outcome, 20 steps. Failure PSNR/SSIM rise 19.51/0.7461 to 25.92/0.8290; success metrics remain nearly unchanged.
Go to primary source ↓e-failed-imagePDF p. 8, Figure 5 and caption
Under the same bad-action condition, the success-only prediction shows a successful grasp in the dotted box, whereas failure-trained prediction resembles the observed failed outcome.
Go to primary source ↓e-scalingPDF p. 8, failure-data scaling paragraph and Figure 6
Three-task average rises 32.7 to 57.3 across failure fractions; the full failure set is about 45% of training data. Put Bottles Dustbin declines at the intermediate fraction.
Go to primary source ↓e-scoringPDF p. 8, value-guided candidate scoring paragraph and Figure 7
Same-checkpoint candidate sweep measures relative completion and latency. The plotted completion peaks at four then falls at eight, despite prose describing smaller gains for larger sets.
Go to primary source ↓e-valuePDF p. 8, Figure 8; p. 17, Appendices G–H and Figure 12; p. 19, Figure 13
Example values fall on missed grasps and recover; a 3×3 placement diagnostic scores the successful center highest. Appendix H explicitly states predictions are unclipped.
Go to primary source ↓e-limitsPDF p. 9, Sections 5–6
Broader data and improved progress estimators are future work; scoring costs an extra pass; online corrections and reinforcement learning are proposed extensions.
Go to primary source ↓e-hardwarePDF p. 15, Appendix C; p. 16, Figure 11
Two YAM arms, D435 main camera, two D405 wrist cameras and GELLO; packed multiview canvas and precomputed VAE latents; baselines share success fine-tuning demonstrations.
Go to primary source ↓e-promptsPDF p. 16, Table 5 and caption
Lists the five real-world task prompts and states that T5 embeddings feed cross-attention.
Go to primary source ↓e-latencyPDF p. 17, Appendix E; p. 19, Table 7, model/inference-time column
RTX PRO 6000 measurements: FACT with failures 380 ms, Motus 1220, Cosmos 620, π₀.₅ 47, π₀ 45; Appendix E identifies FACT's action-only deployment.
Go to primary source ↓8.5 Primary sources
FACT: Failure-Aware Causal Training for World-Action Models ↗
PDF · 7,737 extracted words
Source fingerprint
62b041bdc4a23d4d192a1016e9fa488dd89eba9619a503f5c7a99a30e02596b8