EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies
1. Paper overview
In one sentence: EventVLA forecasts when to save actual camera observations, improving transient-evidence manipulation through sparse memory while adding latency and retaining a fixed-capacity bottleneck. e-probleme-architecturee-robomeme-real-confige-latencye-limit
| At a glance | What to know |
|---|---|
| Research problem | Source description A closed cover can erase the only visible evidence of an object’s identity. Initial frames retain scene layout and recent frames retain motion context, but neither necessarily contains the brief intervening reveal. The problem is choosing which observations to preserve before their downstream relevance becomes immediate. e-problem |
| Core mechanism | Source description A unified VLA shares action-related hidden states with a Keyframe Evidence Memory head, combining deterministic anchors and learned event selection. e-architecture |
| A key reported result | RoboTwin-MeM intermediate-evidence retention: 75.2% (VA+KEM) Reported average task success rate. Eight simulation tasks; QwenOFT backbone. Evaluation trial counts, seeds and split membership are not explicitly specified. 18.0% anchors only; 10.8% MemoryVLA/QwenOFT. Differences from reported averages: +57.2 and +64.4 percentage points. Strong support for event memory here, without isolating forecasting from other KEM training and write-management choices. e-robomeme-benchmarke-sim-config |
| Reading caution | Author claim The authors identify saturation and early-cue eviction for dense-event tasks longer than roughly ten minutes. Five-frame FIFO storage is bounded working memory. e-limite-sim-config |
Core contributions
- Source description
A unified VLA shares action-related hidden states with a Keyframe Evidence Memory head, combining deterministic anchors and learned event selection. e-architecture
- Source description
RoboTwin-MeM introduces eight bimanual tasks involving transient recognition, counting and route imitation. Table 4 lists fifty episodes per task, mean lengths of 430–1544 steps and one to five required keyframes; it does not identify train/test membership. e-benchmark
Figure 2. A predicted event time becomes an image-memory write after the corresponding action step. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the instruction and the two memory components at the top. Visual Anchors keep initial and recent context; Keyframe Event stores selected observations from earlier interactions. Below the Unified VLA Policy, gray prefix tokens are distinguished from orange action tokens. The action-token stream branches toward the Action Head and the Keyframe Head. Follow the orange path from the latter through the probability curve to Future Delayed Commit, then up the right edge into event memory. This is a temporal feedback loop: the predicted peak schedules an observation to be saved after execution reaches that step. Equation (3) defines the curve as keyframe probabilities, not generated imagery. e-architecturee-writee-annotation
What it supports. The architecture makes memory selection aware of action-related hidden states while retaining original observations as context. The action head drives the robot; the keyframe head determines which later camera observations enter the event buffer. Shared hidden states support the unified-policy interpretation, independently of the separate offline model used to annotate demonstrations.
Where the evidence stops. Figure 2 abstracts away the current-observation input and write filtering. Equation (4) explicitly includes the current observation; Appendix A.2 adds thresholding, NMS and cooldown. The illustrated forecast is schematic and supplies no measured prediction accuracy.
2. Motivation
2.1 The problem and the proposed response
A closed cover can erase the only visible evidence of an object’s identity. Initial frames retain scene layout and recent frames retain motion context, but neither necessarily contains the brief intervening reveal. The problem is choosing which observations to preserve before their downstream relevance becomes immediate. e-problem
2.2 What this reading follows
A robot can look directly at a hidden object and still fail later if it forgets what the camera briefly revealed. EventVLA addresses that gap by combining initial and recent images with a small store of interaction events. Its unusual prediction target is the timing of useful future observations: a parallel head schedules memory writes across the upcoming action chunk, then the controller stores the real images as execution proceeds. The evidence is strongest when fixed anchors and learned event capture are compared on RoboTwin-MeM. Read the physical results separately, because those experiments use a different backbone, and read the inference table as part of the method’s tradeoff. e-probleme-architecturee-robomeme-real-confige-latencye-limit
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 | VLA |
| Architecture | One Model |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Conflicts with the recorded classification
VLA and One Model are supported by shared policy/action/KEM architecture; the large annotator is offline. Future-representation-assisted VLA is questionable: KEM predicts event probabilities without future visual representations or dynamics rollout. Autoregressive transformer states are described for simulation, but do not establish autoregressive action decoding for the pi0.5 physical variant. Not-applicable prediction paradigm and quadrant fit the absence of world prediction. e-architecturee-annotatione-real-config
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 remembering the scene from remembering an interaction
Consider Cover Blocks Hard. The initial image records where the covers are, but the later instruction depends on colors revealed as the robot opens and recloses them. A recent-history window helps describe arm motion and task progress; once it slides beyond a reveal, that color evidence is lost again. EventVLA therefore gives different jobs to anchors and event memory. The anchor component keeps predictable context, while KEM selects observations whose usefulness may outlast their visibility. This interpretation explains why the two benchmark results should be read together: anchors alone work well on RMBench, yet improve substantially when KEM is added on RoboTwin-MeM. It remains a task-specific inference, however, rather than proof that all long-horizon manipulation requires the authors’ chosen event representation. e-probleme-architecturee-benchmarke-rmbe-robomem
Figure 3. The benchmark places information between the initial scene and the moment when it is needed. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each short sequence from left to right and use the blue borders to locate the evidence the authors designate as task critical. Pick the Unhidden Block requires remembering the colors revealed beneath covers; Cover Blocks Hard later reopens covers in a specified color order. Reproduce Route instead asks the policy to retain a randomly demonstrated spatial sequence. The label n counts required intermediate keyframes, rather than actions or video length. Two tasks explicitly have ranges: seal finding needs one to four keyframes and button pressing two to five. Table 4 supplies the corresponding instructions and episode statistics behind these visual examples. e-benchmarke-robomeme-latency
What it supports. The tasks distinguish remembering an initial arrangement from preserving information created by an interaction. Once a cover closes, recent images may again show indistinguishable covers. This gives the memory module a specific role: retaining a reveal, displacement or sequence event until a subsequent decision needs it.
Where the evidence stops. The highlighted frames express the authors’ task decomposition; the picture alone does not prove each frame is causally necessary. Tables 2 and 9 display fixed n headings for tasks whose actual ranges are variable in Figure 3 and Table 4.
5.2 Teach a schedule with temporally tolerant supervision
The training target is an event timestamp supplied by the offline annotation pipeline. Because several nearby frames may all reveal the same useful fact, Appendix A.1 spreads a timestamp into raised-cosine soft labels. The KEM head then learns one probability per future action step using binary cross-entropy, alongside the action objective. Early training constructs memory from annotations frequently; the teacher-to-student schedule gradually makes the policy experience its own predicted writes. At inference, a high score is only a candidate: a local-maximum test and a strict cooldown reduce repeated captures of one event. These mechanisms address different problems—label tolerance, exposure to imperfect memory, and write redundancy. The hard-label and no-NMS ablations test parts of that design, but the paper leaves the numerical smoothing radius unresolved. e-traininge-writee-annotatione-ablation
5.3 Follow the prediction all the way to a real memory update
A future keyframe score is not an image and cannot itself reveal the object under a cover. The robot must execute the action, acquire the selected observation and commit that image into its bounded event buffer. Only a later policy call can use the stored visual evidence as context. This distinction separates the learned prediction from its physical effect and prevents interpreting KEM as a learned visual world simulator. It also exposes practical dependencies: capture times must align with execution, duplicate writes must be suppressed, and useful early frames must survive FIFO eviction. The reported buffer has capacity five, while full-model profiling takes 1.09 seconds per chunk on average. My reading is that capture reliability, retention and scheduling latency are inseparable parts of evaluating this memory mechanism. e-architecturee-writee-sim-confige-latencye-limit
5.4 Training and inference
During training
Offline annotations become raised-cosine soft targets around event times. Horizon-averaged binary cross-entropy trains KEM jointly with the action loss, weighted by 0.1. Teacher-forcing probability decays linearly from one to zero, transferring memory construction from annotations to predictions. e-traininge-sim-config
Simulation fine-tunes QwenOFT/Qwen3-VL-4B-Instruct end to end for 80,000 AdamW steps, with 224×224 images, backbone/head learning rates 1e-5/1e-4 and fifty-step actions. RMBench trains anchors only. Physical experiments adapt pi0.5/PaliGemma: 32-dimensional actions, 60,000 steps, batch 32 and peak learning rate 5e-5. No frozen module is identified. e-sim-confige-real-config
During inference
Full KEM uses threshold 0.55, NMS radius eight, cooldown ten steps and capacity five. Eq. (8) retains local maxima at or above threshold; cooldown requires a gap strictly greater than ten. Physical anchors use t−60, t−40 and t−20 alongside the initial frame. The annotator operates offline. e-writee-sim-confige-real-confige-annotation
5.5 Implementation flow
- Assemble explicit visual context
Visual anchors and FIFO event frames form the memory. Although Eq. (2) introduces a general recent window, simulation uses initial, t−30 and t−15 frames. Temporally ordered historical images and the current observation enter the vision encoder, preserving image evidence instead of one compressed latent memory. e-architecturee-sim-config
- Forecast the write schedule
KEM applies an MLP and sigmoid to final action-related transformer states. Each output scores an upcoming execution step as a potential keyframe. Figure 2 routes action tokens to both action and keyframe heads; the forecast concerns event timing, not future pixels. e-architecture
- Commit observations during execution
The forecast schedules delayed capture of the actual image. Thresholding, local-maximum suppression and cooldown reduce duplicate writes; FIFO bounds event storage. Subsequent actions read the saved observations, closing the loop through visual feedback. e-architecturee-write
6. Experiments & results
EventVLA augments a VLA controller with sparse images of earlier events. A parallel head forecasts which upcoming action steps merit a memory write; the system saves the actual image when that step occurs. Its diagnostic RoboTwin-MeM result rises from 18.0% with fixed anchors to 75.2% with KEM. This supports memory-assisted robot control, with no future-image prediction objective.
6.1 Read the original evidence
Table 1. On RMBench, the reported gain comes from visual anchors alone. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the right column as a reported average success percentage, with higher values indicating more completed tasks. The final block is the most direct architectural comparison: full visual anchors, removal of the initial observation, and removal of recent history. These rows test two types of historical context without activating KEM. The earlier blocks broaden the comparison to reactive policies and other memory approaches. In particular, distinguish MemoryVLA based on OpenVLA from the QwenOFT reimplementation, since EventVLA’s simulation backbone is also QwenOFT. Table 8 expands this average into nine individual tasks, making clear that a strong aggregate need not imply every memory problem is solved. e-rmbe-sim-config
What it supports. Visual anchors report 67.8%, compared with 33.7% without the initial frame and 23.8% without recent history. Both components contribute within this setup. Mem-0 reports 42.0% and MemoryVLA/QwenOFT 41.7%, but the central lesson for KEM is a boundary: these RMBench gains do not require learned event capture.
Where the evidence stops. No uncertainty is reported. Table 8 still gives only 3% on Press Button. Its displayed task rates average about 67.9%, while Tables 1 and 8 report 67.8%; the underlying precision or aggregation needed to resolve that discrepancy is absent.
Figure 4. The physical evaluation measures completed robot tasks with a pi0.5-based EventVLA. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the displayed panel order: Find Block Easy, Pick-X-Times, Find Block Hard and Pick in Order. Each panel aligns a success chart with a photograph and an instruction. Gray denotes reactive pi0.5, blue the reimplemented piMEM baseline, and green EventVLA. The green arrows show absolute differences in percentage points from the blue bars. Distinguish the two hidden-block settings: Easy starts by covering a known block, whereas Hard requires searching under cups and remembering the successful reveal. Section 5.3 specifies twenty independent trials per task. Appendix B.3 also matters here: the physical model uses pi0.5/PaliGemma rather than the QwenOFT simulation backbone. e-reale-real-confige-real-discrepancy
What it supports. In this figure’s order, EventVLA reports 90%, 90%, 60% and 75%, versus piMEM’s 50%, 30%, 30% and 40%. These are physical task-completion measurements. The counting task’s difference is sixty percentage points; it is not a sixty-percent relative improvement or a result from generated video.
Where the evidence stops. Twenty trials per task and no confidence intervals limit precision. The introduction says success reaches up to 80%, conflicting with two 90% bars and Section 5.3. This edition preserves the detailed measurements and discloses the inconsistency.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin-MeM intermediate-evidence retention Eight simulation tasks; QwenOFT backbone. Evaluation trial counts, seeds and split membership are not explicitly specified. | 75.2% (VA+KEM) Reported average task success rate | 18.0% anchors only; 10.8% MemoryVLA/QwenOFT. Differences from reported averages: +57.2 and +64.4 percentage points. Strong support for event memory here, without isolating forecasting from other KEM training and write-management choices. e-robomeme-benchmarke-sim-config |
| RMBench visual anchoring Nine simulation tasks; EventVLA uses anchors only. | 67.8% Reported average task success rate | Mem-0: 42.0%; MemoryVLA/QwenOFT: 41.7%; no initial anchor: 33.7%; no short-term anchors: 23.8%. Supports initial/recent observations, without evaluating KEM. Press Button remains only 3% in the per-task breakdown. e-rmb |
| RoboTwin-MeM KEM ablations Eight-task benchmark; listed variants versus full VA+KEM. | Implicit memory: 24.9%; hard labels: 48.8%; no NMS: 53.4%; capacity two: 32.0%; chunk thirty: 31.1%; chunk fifteen: 13.6%. Reported average task success rate | Full configuration: 75.2%. Representation, supervision and buffer choices matter. Shorter action chunks also change control timing, confounding a foresight-only explanation. e-ablation |
| RoboTwin 2.0 reactive manipulation Standard Easy/Hard settings; split and trial details are not expanded here. | Easy 83.8%; Hard 81.6%. Task success rate | QwenOFT: 80.0% and 78.0%; QwenFast reaches 83.2% on Hard. Improves over QwenOFT without being the strongest listed Hard result. e-markov |
| Real-world memory tasks on ARX ACONE Physical pi0.5-based EventVLA; twenty independent trials per task. Figure order: Find Block Easy, Pick-X-Times, Find Block Hard, Pick in Order. | 90%, 90%, 60%, 75%, respectively. Task success rate | Reimplemented piMEM: 50%, 30%, 30%, 40%; reactive pi0.5: 10%, 0%, 10%, 0%. Executed-robot evidence, without reported confidence intervals. Figure 4 and Section 5.3 support these values despite the introduction’s conflicting 80% maximum. e-reale-real-confige-real-discrepancy |
| RoboTwin-MeM inference cost Simulation chunk-generation profiling; hardware unspecified. | VA+KEM: 1.09 s/chunk; 0.94 chunks/s. Mean seconds per chunk; mean chunks per second | Anchors only: 0.96; 1.07. QwenOFT: 0.36; 2.91. Memory adds computational cost. Chunk throughput is not low-level control frequency or a universal real-time guarantee. e-latency |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. Transient-event tasks expose the difference between fixed anchors and learned memory writes. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the adjacent EventVLA VA-only and VA+KEM rows, then move across tasks before reading Total average. This keeps the memory comparison within the EventVLA family and reveals uneven gains: Rearrange Blocks Hard remains 62% in both rows, while Reproduce Route changes from 0% to 98%. The gray rows below the horizontal divider are design ablations, each to be compared with the full VA+KEM row. They alter representation, supervision, write filtering, storage capacity or action horizon. Bold and underlined entries mark the source’s best and second-best main-comparison results; the gray ablations also need to be read numerically rather than ranked from typography alone. e-robomeme-ablatione-sim-config
What it supports. The reported average increases from 18.0% to 75.2%, a 57.2-percentage-point difference using the stated totals. Removing NMS gives 53.4%, and reducing capacity to two gives 32.0%. Together, these comparisons support a role for retaining diverse event images rather than merely increasing the amount of recent context.
Where the evidence stops. The rows support the complete tested design but do not isolate every causal explanation. Shorter chunks change both action execution and the forecast window. The paper provides no matched-budget reactive keyframe classifier or uncertainty estimates to settle whether foresight itself accounts for the gain.
Table 9. The memory improvement has a measurable cost in chunk-generation latency. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each model as a pair of rows. The caption defines Latency as seconds per generated chunk and Throughput as generated chunks per second, so smaller latency and larger throughput are desirable. Compare the Total average column first, then inspect task differences. QwenOFT is the reactive reference; anchors only adds historical visual input; VA+KEM adds dynamic event memory. This decomposition helps distinguish the cost already paid for static context from the incremental cost of event frames. Keep the units tied to whole chunks: the table does not measure individual actuator commands, controller update frequency or end-to-end observation-to-motion delay. e-latency
What it supports. Average latency rises from 0.36 seconds for QwenOFT to 0.96 for anchors and 1.09 for full EventVLA. The incremental KEM difference is 0.13 seconds using reported averages, while full throughput is 0.94 chunks per second. Memory quality therefore needs to be considered alongside a substantial cost relative to the reactive baseline.
Where the evidence stops. Profiling hardware and execution synchronization are unspecified, and separately averaged latency and throughput need not be exact reciprocals. The authors’ claim of practical real-time suitability is not a guarantee for arbitrary controllers, event densities or robot hardware.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify saturation and early-cue eviction for dense-event tasks longer than roughly ten minutes. Five-frame FIFO storage is bounded working memory. e-limite-sim-config
No uncertainty estimates accompany the main success tables. Table 8’s displayed integer task rates average about 67.9%, whereas its total reports 67.8%; this report preserves the stated total because underlying precision or aggregation is unspecified. Qualitative successes do not establish failure prevalence. e-rmbe-robomeme-reale-qualitative
The introduction’s real-world maximum conflicts with detailed results. Fixed n headings in Tables 2 and 9 also compress Table 4’s variable ranges for seal finding and button pressing; they are not universal per-episode counts. e-real-discrepancye-benchmarke-robomeme-latency
7.2 Questions for discussion
- Would forecasting improve success with action timing and image-token budgets held constant?
- When several events occur within the cooldown, which missed observation determines downstream failure?
8. Reproducibility audit
8.1 Requirements and known gaps
Annotation uses Qwen3-VL-235B-A22B-Instruct-FP8, vLLM and eight A800 GPUs, plus required keyframe counts and few-shot examples. Reported timing errors are below ten steps on average in simulation and within fifty steps physically; validation sample counts are absent. e-annotation
Resolve the unspecified smoothing radius, overlapping-label handling, NMS ties, timestamp synchronization and early-history padding. Record software versions, simulation splits, seeds, real demonstration counts and training/profiling hardware. Table 4’s episode totals and the annotation server do not establish these missing protocol details. e-traininge-writee-benchmarke-sim-confige-real-confige-latency
Proposed checks should compare memory schedules at fixed action timing and measure duplicate-write eviction. The illustrated edition specifies controls and falsifiable observations; neither check has been run. e-ablatione-write
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 forecasting help when action execution is held fixed?
Reader-proposed check, not performed: on Cover Blocks Hard and Reproduce Route, compare the full KEM schedule with a matched current-frame event classifier and a uniformly spaced write schedule. Keep the action backbone, fifty-step action chunk, demonstrations, visual anchors, maximum five event frames and training budget fixed; match the mean number of writes. Use held-out layouts and multiple training seeds. Measure task success, capture of annotated evidence windows and the time between each reveal and its first later use. A forecasting advantage should persist at matched write budgets and should coincide with capturing transient events inside the action chunk. If success differences vanish, the shorter-chunk ablation does not by itself demonstrate a distinct benefit from foresight. e-benchmarke-architecturee-ablatione-sim-confige-training
Check 2: Do duplicate writes cause failure by evicting early evidence?
Reader-proposed check, not performed: compare the reported NMS/cooldown pipeline, no NMS with capacity five, and no NMS with enough capacity to retain every distinct annotated event on the same Cover Blocks Hard episodes. Keep the action model, threshold, training setup and evaluation seeds fixed. Log selected timestamps, duplicate captures, FIFO eviction times, survival of each color reveal until reopening, and final success. Include an oracle distinct-event buffer as a diagnostic upper control, explicitly labeled as using privileged annotations. The proposed eviction mechanism predicts that the no-NMS condition loses early reveals and that extra capacity or oracle deduplication rescues both retention and success. Persistent failure despite retained evidence would point toward retrieval or action errors. e-ablatione-writee-benchmarke-sim-confige-limit
8.3 Reading coverage
Visual audit: Visually inspected the title/authors/version and affiliations; Figures 1–8; Tables 1–9; method equations; training and annotation pages; both simulation and physical configuration tables; and qualitative appendix rollouts. All six final crops were viewed, with two table crops tightened and re-inspected. Figure 2’s action-token branches and delayed-commit return were checked against Eqs. (3)–(4), and write inequalities against Appendix A.2. No arrow-direction conflict was found. The abstract’s broad improvement statement is not used as a pooled result. The introduction/Figure 4 numerical conflict, Table 8 aggregation discrepancy and variable keyframe counts are disclosed in the report. Reference pages 9–12 were read in the complete text but not visually inspected. No external supplements, code or videos were inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23. 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 page and Abstract
- 1 Introduction
- 2 Related Work
- 2.1 Memory-Augmented Policies for Long-Horizon Manipulation
- 2.2 Memory-Oriented Manipulation Benchmarks
- 3 EventVLA Framework
- 3.1 Problem Formulation and Foundational Visual Anchors
- 3.2 Keyframe Evidence Memory (KEM) Module
- 3.3 End-to-End Training and Inference Details
- 4 RoboTwin-MeM Benchmark
- 5 Experiments
- 5.1 Performance on Simulation Benchmarks
- 5.2 Ablation Analysis of EventVLA
- 5.3 Real-World Robot Evaluation
- 6 Limitations
- 7 Conclusion
- Acknowledgments
- References
- Appendix
- A Implementation Details of EventVLA
- A.1 Training Formulations and Curriculum Strategy
- A.2 Online Inference and Post-Processing
- A.3 Automated Keyframe Annotation Pipeline
- B Experimental Setups and Benchmarks
- B.1 RoboTwin-MeM Benchmark Details
- B.2 Real-world Tasks Details
- B.3 Network Architecture and Hyper-parameters
- C Extended Experimental Results and Analysis
- C.1 Detailed Per-Task Breakdown on RMBench
- C.2 Extended Ablation Analysis and Inference Efficiency
- D Qualitative Visualizations
- D.1 Simulation Rollouts in RoboTwin-MeM
- D.2 Real-World Robot Execution Sequences
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.
- Reviewed the supplied 23-page arXiv:2606.20092v2 dated 29 June 2026. The catalog records submission on 18 June 2026; the exact title and thirteen-author list match. No earlier revision was supplied, so cross-version scientific differences cannot be assessed.
- Separate supplemental material availability has not been fully verified; no separate supplement was supplied.
- The acquisition text omitted figure imagery. This was resolved by inspecting the retained PDF: all eight figures and all nine tables were visually reviewed. Reference pages 9–12 were read as text but not rendered.
- Code, models, datasets and external project links were not inspected; the paper’s availability statement remains unverified. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author/affiliation block and arXiv margin
The exact catalog title and all thirteen authors appear on the title page. The margin identifies arXiv:2606.20092v2, cs.CV, 29 June 2026. Eight institutions are listed.
Go to primary source ↓e-problemPDF pp. 2–3, Introduction and Section 3.1, Eq. (1)
Initial scene information and recent motion cues can be insufficient when intermediate interactions briefly expose information later hidden. The policy conditions actions on the current observation, historical memory and instruction.
Go to primary source ↓e-architecturePDF p. 4, Sections 3.1–3.2, Figure 2 and Eqs. (2)–(3); p. 5, Eq. (4)
Visual anchors and event frames form an explicit image memory. An MLP with sigmoid maps final transformer action-related hidden states to future keyframe probabilities. Delayed writes store actual observations; anchors, event frames and current observation enter the vision encoder together.
Go to primary source ↓e-trainingPDF p. 5, Section 3.3; p. 13, Appendix A.1, Eqs. (6)–(7)
Raised-cosine targets around annotated event timestamps supervise horizon-averaged binary cross-entropy. The joint action/KEM objective is trained with a teacher-forcing probability that decays linearly from one to zero. The dilation radius R is defined without a numerical setting.
Go to primary source ↓e-writePDF p. 13, Appendix A.2, Eq. (8); p. 14, opening paragraph; p. 5, FIFO description
Candidate steps satisfy the confidence threshold and equality to the local probability maximum. Sequential commits additionally require elapsed time strictly greater than the cooldown. The event buffer uses bounded FIFO eviction.
Go to primary source ↓e-annotationPDF p. 14, Appendix A.3, preprocessing, prompt box and reliability paragraph
Offline annotation uses Qwen3-VL-235B-A22B-Instruct-FP8 with vLLM on eight NVIDIA A800 GPUs. Sampled multiview frames, few-shot examples and a required keyframe count guide timestamp annotation. The authors report average absolute simulation error below ten timesteps and real-world error within fifty timesteps, without a validation sample count or error distribution.
Go to primary source ↓e-benchmarkPDF p. 5, Figure 3; p. 6, Section 4; p. 15, Table 4 and Appendix B.1
RoboTwin-MeM builds on RoboTwin 2.0/SAPIEN. Table 4 lists eight tasks, fifty episodes each, mean lengths 430–1544 steps and required intermediate keyframes from one to five. Find Seal and Seal Stamp has range [1,4]; Press Button Keyframe has [2,5]. Table 4 does not assign its episode counts to train/test splits.
Go to primary source ↓e-sim-configPDF p. 17, Appendix B.3 and Table 5; p. 18, Table 6
Simulation uses QwenOFT/Qwen3-VL-4B-Instruct, 224-by-224 images, fifty-step actions, AdamW for 80,000 steps and learning rates 1e-5 for the backbone and 1e-4 for heads. Anchors are initial, t−30 and t−15 frames. Full KEM uses threshold 0.55, capacity five, NMS radius eight, cooldown ten and loss weight 0.1.
Go to primary source ↓e-real-configPDF p. 17, Appendix B.3, final paragraph; p. 18, Table 7
Real-robot training adapts pi0.5/PaliGemma with 32-dimensional actions over fifty steps. Joint AdamW training lasts 60,000 steps, global batch size 32, bfloat16, peak learning rate 5e-5 and 2,000 warm-up steps. Anchors are initial, t−60, t−40 and t−20 frames; full KEM settings match Table 6.
Go to primary source ↓e-rmbPDF p. 6, Table 1 and Section 5.1; p. 19, Table 8
RMBench visual-anchors-only EventVLA reports 67.8% overall, against Mem-0 42.0% and MemoryVLA/QwenOFT 41.7%. Removing the initial or short-term anchors gives 33.7% or 23.8%. Table 8 displays nine task columns, including 3% for Press Button.
Go to primary source ↓e-robomemPDF p. 7, Table 2, baseline and EventVLA rows
RoboTwin-MeM reports total averages of 75.2% for VA+KEM, 18.0% for VA only and 10.8% for MemoryVLA/QwenOFT. VA+KEM per-task values are 62, 93, 90, 54, 94, 63, 98 and 48 percent, in displayed column order.
Go to primary source ↓e-ablationPDF p. 7, Table 2, gray ablation rows and Section 5.2; p. 8, opening continuation; pp. 19–20, Appendix C.2
Reported averages are 24.9% for implicit memory, 48.8% for hard labels, 53.4% without NMS, 32.0% for capacity two, 31.1% for chunk size thirty and 13.6% for chunk size fifteen. The authors attribute these changes to compression, label sparsity, redundant writes, eviction and reduced foresight.
Go to primary source ↓e-markovPDF p. 7, Table 3 and standard Markovian evaluation paragraph
On RoboTwin 2.0 Easy/Hard, EventVLA reports 83.8%/81.6% and QwenOFT 80.0%/78.0%. QwenFast reports 83.2% on Hard, above EventVLA.
Go to primary source ↓e-realPDF p. 8, Figure 4 and Section 5.3
Four tasks on ARX ACONE are each evaluated over twenty independent trials. In figure order, EventVLA obtains 90%, 90%, 60%, 75%; piMEM obtains 50%, 30%, 30%, 40%; pi0.5 obtains 10%, 0%, 10%, 0%. Figure 4 orders Pick-X-Times before Find Block Hard.
Go to primary source ↓e-real-discrepancyPDF p. 3, Introduction, opening paragraph; p. 8, Figure 4 and Section 5.3
The introductory claim says real-world success reaches up to 80%, while the detailed real-world figure and results paragraph report two tasks at 90%.
Go to primary source ↓e-limitPDF p. 8, Section 6, Limitations
The authors identify saturation and premature eviction in a bounded event buffer as a limitation for dense-event, exceptionally long tasks, giving more than ten minutes as an example.
Go to primary source ↓e-latencyPDF p. 20, Table 9 and Appendix C.2, Inference Efficiency
Mean latency/throughput are 0.36 seconds per chunk/2.91 chunks per second for QwenOFT, 0.96/1.07 for anchors only, and 1.09/0.94 for VA+KEM. These are chunk-generation measurements; the section does not specify profiling hardware, low-level control frequency or synchronization policy.
Go to primary source ↓e-qualitativePDF p. 16, Figure 5 and Appendix B.2; pp. 20–23, Appendix D and Figures 6–8
Selected simulation and physical execution sequences illustrate hidden-object inspection, counting, spatial routes and order following. Figure 5 marks keyframes with blue borders; Figures 6–8 provide longer successful rollouts without aggregate failure statistics.
Go to primary source ↓8.5 Primary sources
EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies ↗
PDF · 10,085 extracted words
Source fingerprint
e09a7d8a095e0948759c8e9c82922b7c04c031a6b9fb06c0bcef99fc2ea983b3