EWAM: An Enhanced World Action Model for Closed-Loop Online Adaptation in Embodied Intelligence
1. Paper overview
In one sentence: EWAM trades extra inference work for more efficient simulated execution by adding memory, anomaly-aware correction and filtered online adaptation to a frozen WAM. e02e05e08e09e10e15e17e19
| At a glance | What to know |
|---|---|
| Research problem | Source description A plausible action can still collide, miss a grasp or waste motion when realized dynamics depart from predictions. EWAM asks whether a fixed pretrained WAM can exploit execution feedback without collecting another target-task demonstration dataset. Its zero-shot label permits memory growth and lightweight online learning. e01e02 |
| Core mechanism | |
| A key reported result | BananaInBowlTask execution efficiency: EWAM: 100%; 9.27±0.94 s; 0.83±0.08 m. Success; completion time; end-effector path. RoboLab; same frozen backbone/interface; position σ=0.02 m and lighting ±15%; 5 seeds×25 trials; mean±95% CI. Frozen baseline: 100%; 25.60±2.31 s; 1.81±0.12 m. Reported time/path reductions are 63.8%/54.1%. Success is saturated; the finding concerns execution quality. Exact paired-test values are not supplied. e09e10 |
| Reading caution | Source description No physical-robot or sim-to-real result is reported. Recovery depends on simulator checkpoints, interfaces are backbone-specific, and compact memory remains vulnerable to occlusion, novel dynamics, ambiguous objects and long-horizon errors. e15 |
Core contributions
Figure 1. Execution feedback returns through filtering, memory and lightweight adaptation. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the input column, then follow the gray backbone toward the environment on the right. The blue blocks describe experience retrieval, strategy selection, mismatch handling and action refinement; the lower loop shows how execution outcomes can affect later behavior. Solid arrows denote data flow and dashed arrows feedback or updates. Treat this as a conceptual overview: the graphic places strategy selection before mismatch detection and draws a direct backbone-to-output arrow. Equations (1)–(10) instead specify VLE/AR processing, memory inside the DiT, state/action heads, anomaly detection, routing and correction before execution. The equation-defined sequence supports the mechanism described here. e02e03e07e19
What it supports. The architectural contribution is a feedback path around a frozen policy, with memory also entering its internal representation. An executed action produces evidence that may alter future context or lightweight parameters. This differs from merely predicting a plausible future: the environment supplies the diagnostic signal that closes the loop.
Where the evidence stops. Figure 1 also places self-learning before memory writing; Section 3.7 and Appendix B specify admission before scheduled updates. The schematic does not establish the exact computation graph, and the frozen-backbone constraint does not mean the inserted modules are frozen.
2. Motivation
2.1 The problem and the proposed response
A plausible action can still collide, miss a grasp or waste motion when realized dynamics depart from predictions. EWAM asks whether a fixed pretrained WAM can exploit execution feedback without collecting another target-task demonstration dataset. Its zero-shot label permits memory growth and lightweight online learning. e01e02
2.2 What this reading follows
A robot can eventually finish a task while repeatedly colliding, missing grasps or taking an unnecessarily long path. EWAM treats those execution details as useful feedback. A frozen Cosmos3-Nano--Policy-DROID backbone supplies state and action candidates; added modules retrieve experience, diagnose risk and choose how to intervene. Successful execution alone does not qualify a trajectory for learning: it must also pass a safety and efficiency gate. The paper reports large local time reductions, but the experiment includes accumulated online experience and simulator rollback. Read the tables alongside the protocol and schematic caveats, because several figure labels disagree with the formal method or evaluation description. e02e05e08e09e10e15e17e19
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
The shared DiT representation feeds both state and action heads, supporting One Model at the backbone level; EWAM inserts memory into that path rather than merely co-training separate predictors. Its added mechanism is feedback routing, correction and filtered adaptation, consistent with Other mechanisms/Outside quadrants. Memory and policy post-training are supported; the stated objectives are supervised, so the combined WM-RL category should not be read as evidence of an RL algorithm. e02e03e05e06
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 candidate from the evidence that judges it
The frozen backbone first turns vision, proprioception and language into a context-conditioned DiT representation. That representation feeds both a state prediction and a raw action chunk. EWAM then asks whether executing that candidate looks risky. At this moment it can use the previous prediction residual, disagreement with a candidate-action dynamics prediction and learned risk scores. It cannot yet compare the candidate's predicted next state with the next observation. That comparison becomes possible only after execution. This timing distinction explains why the paper separates decision-time anomaly features from the realized residual stored in experience. A discrete router selects unchanged execution, bounded correction or rollback recovery. For implementation, follow Section 3.4 and the explicit branch definition in Equation (25); the later gated-system notation compresses the temporal sequence and leaves a realized-error trigger ahead of the transition that supplies it. e02e04e05e07
Figure 3. A completed trajectory must pass every quality criterion before it can teach the policy. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the six checks as a conjunction, not as a vote: one failed criterion sends the trajectory down the red branch. The green branch writes qualified experience into memory and makes it eligible for online updates. Equation (52), on the preceding page, gives the reported operating rule: minimum safety and task scores, maximum gripper force, an EE SPARC upper bound, and limits on path length and time. The diagram's smoothness symbol instead uses a lower-bound inequality, and its force condition is an interval. The source supplies no transformation connecting those generic figure conditions to the equation, so the numerical rule should be quoted from Equation (52). e08e02e06e20
What it supports. Filtering controls which deployment experience becomes future supervision. A successful but inefficient trajectory can still be rejected, leaving it available for diagnostics without adding it to memory or replay. That makes admission policy part of the adaptation mechanism, rather than just a reporting metric applied after learning.
Where the evidence stops. The figure's 'zero modification' claim cannot describe the whole architecture, which inserts neural layers. The six thresholds are task-specific; neither their generality nor their necessity is isolated experimentally. Units/normalization of the force limit and the SPARC implementation remain incomplete.
5.2 Understand what can change while the backbone stays frozen
EWAM has two ways to adapt. First, admitted experience changes the memory database: a later DiT pass can retrieve different context even if every parameter is identical. Second, scheduled training can change the inserted modules and adapters using admitted replay. The backbone remains fixed in both cases. Memory-first therefore describes an ordering of adaptation mechanisms, not an absence of learning. Appendix A resets memory at the start of each seed and lets it grow along the matched trial sequence, so early and late episodes need not use the same effective policy. Offline supervision can already have prepared the added modules, while deployment supplies further simulator-derived targets. The zero-shot promise is specifically that evaluation introduces no additional target-task demonstration dataset. Reproducing that promise requires knowing memory initialization, training data and update scheduling; the supplied paper does not completely specify them. e03e06e07e09e16
5.3 Read efficiency gains without overclaiming their cause
A useful interpretation starts with the saturated success rate on BananaInBowlTask. Faster completion and shorter paths can reveal improvement even when success cannot increase. The ablations then ask which parts of EWAM matter: removing self-learning or correction is much more costly than removing rollback on this particular task. My interpretation is that these aggregate results motivate a learning-curve study rather than proving a universal priority ordering among modules. The no-self-learning condition changes experience accumulation, and the memory-only/update-only comparison changes what information can influence inference. Neither contrast tests an unrestricted filter or measures forgetting on unrelated tasks. Stacking shows a success gain for the full system, but does not identify rollback's separate contribution. Finally, the latency table shows why execution speed and inference speed must remain distinct, and the inconsistent figure captions prevent reliable uncertainty claims from the ablation graphic. e10e11e13e14e17e20
5.4 Training and inference
During training
Offline preparation combines action/state squared errors, memory cosine loss, binary risk and residual supervision, routing cross-entropy and anomaly-conditioned safe-action regression. Their weights are 1.0/0.5/0.3/0.8/0.6/1.2. Safe targets may come from simulator rollback, recovery primitives or admitted successful memory. e05e06
Only inserted layers and adapters receive updates; VLE, AR, DiT and output heads stay frozen. Online replay contains admitted samples only. Table 2 specifies learning rate 10^-4, rank-16 adapters and buffer size 1000; the exact update schedule and offline training corpus are unspecified. e06e16
During inference
Retrieval affects the next generation without changing parameters. Realized next-state error is computed after execution; it is not an observation available for the current candidate. Section 3.4 supplies this causal ordering even though Eqs. (44)–(47) place a realized-error trigger before the transition. e04e07
Equation (52) requires SafetyScore≥0.75, TaskScore≥0.80, maximum gripper force≤0.80, EE SPARC≤−2.5, path≤1.2 m and time≤15 s. Rejected trajectories remain diagnostic-only. These are scenario-specific rules; the force quantity's units/normalization and SPARC computation are not fully specified. e08e16
5.5 Implementation flow
- Encode and generate
VLE encodes observation, proprioception and language; the AR reasoner produces context. The DiT generates a shared hidden representation used by frozen state and action heads. Robot tokens represent camera-pose delta, end-effector-pose delta and gripper state. e02e03
- Condition on admitted experience
At DiT layer 6, memory attention enters through a residual gate of 0.4 using five retrieved experiences. Proposed records store execution context, outcome and rollback anchor. The implemented compact backend is less rich than the full multimodal memory description. e03e15
- Diagnose and route
A six-component vector combines previous prediction error, candidate dynamics inconsistency and learned collision, empty-grasp, hallucination and force-risk scores. Routing combines this vector, AR context and memory. Reported operating regions are direct below 0.3, conservative from 0.3 to below 0.7, and rollback at or above 0.7. e04e05
- Execute, recover and admit
Direct routing keeps the raw action; conservative routing adds bounded correction; rollback restores a stable simulator checkpoint and replans. Without a stable state, recovery replans from the current state. Executed transitions are then filtered before memory writing or training. e05e07
6. Experiments & results
EWAM adds memory-conditioned generation, anomaly detection, discrete routing and action correction to a frozen Cosmos3-Nano--Policy-DROID policy. Its local advantage is execution efficiency: BananaInBowlTask time falls from 25.60 to 9.27 seconds while success stays at 100%. This adaptation uses accumulated simulation experience and lightweight updates, and increases inference latency. The evidence supports a bounded simulation deployment result, with unresolved figure and protocol inconsistencies.
6.1 Read the original evidence
Table 5. Equal success rates conceal large differences in execution time and path length. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the success column: both methods finish every tested BananaInBowlTask episode. Then compare time and path, where smaller values indicate more efficient execution. The caption identifies mean plus or minus 95% confidence intervals over five seeds and 25 trials per seed. Collision, empty-grasp and rollback entries are per-episode event counts, despite nearby prose sometimes calling them rates. The baseline is the same frozen Cosmos3 policy without EWAM layers. Appendix A matters here: models see matched deterministic trial sequences, and EWAM memory resets per seed before growing through admitted experience. These are aggregated online-adaptation results. e09e10e18
What it supports. EWAM reports 9.27±0.94 seconds versus 25.60±2.31, and 0.83±0.08 metres versus 1.81±0.12. The reported relative reductions are 63.8% and 54.1%. Collision and empty-grasp counts also fall, while EWAM uses rollback. The evidence concerns execution efficiency and recovery behavior because task success is already saturated.
Where the evidence stops. The separated time intervals are descriptive evidence, not supplied paired-test results. Exact test values require per-seed logs. The comparison does not measure first-trial performance separately, nor does it show superiority to the other models listed in the capability table.
Table 4. Shorter robot episodes require a somewhat slower policy inference path. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the first three rows as the baseline inference components. The next four rows report incremental costs for the added neural layers; their latency increments sum to 10.2 milliseconds. The two bold total rows provide the end-to-end comparison. Unlike the task-result tables, this table's caption specifies mean plus or minus standard deviation over 100 inference trials. The reported workstation uses two RTX 5880 Ada GPUs with 48 GB each. Keep the throughput column separate from controller frequency: the source labels it step/s but does not fully explain how those steps relate to the latency measurement or action-buffer execution. e13e10
What it supports. EWAM increases measured latency from 66.2±2.4 to 76.4±2.5 milliseconds and reported GPU memory from 1690 to 1820 MB. The largest layer latency increment is memory retrieval at 3.1 milliseconds. Better task efficiency therefore comes from execution behavior, not from reducing the measured neural inference time.
Where the evidence stops. The measured path exceeds the 50 ms budget for a strict 20 Hz loop. Section 4.3 discusses buffered actions or asynchronous decoupling; the 520 step/s throughput entry should not be read as evidence of a 520 Hz end-to-end robot controller.
Table 8. The benefit extends to two additional simulation tasks, with a success gain on stacking. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each task's pair of rows together before comparing different tasks. The two container tasks already have 100% success for both policies, so their improvements are in time, path and collision counts. Stacking adds a success difference because the frozen baseline does not saturate that metric. The caption reports 95% confidence intervals over five seeds and 25 trials per task, making 375 evaluations per model across the three tasks. Position and lighting perturbations are shared across conditions. The table establishes broader coverage than the single banana-in-bowl test, but all rows still belong to the same RoboLab evaluation environment. e09e14e15
What it supports. On BananasInBinOneMoreTask, time falls from 18.40 to 11.20 seconds with success unchanged. On BlockStackingOrderAgnosticTask, success rises from 87.2% to 94.4%, a 7.2-percentage-point gain, while time falls from 42.30 to 28.60 seconds. These paired results support local transfer of the complete EWAM system.
Where the evidence stops. This table does not isolate rollback as the cause of improved stacking success. It also does not validate new robot embodiments, physical recovery or broad open-world generalization. The compact memory and scenario-specific filtering rules remain important limits on that interpretation.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| BananaInBowlTask execution efficiency RoboLab; same frozen backbone/interface; position σ=0.02 m and lighting ±15%; 5 seeds×25 trials; mean±95% CI. | EWAM: 100%; 9.27±0.94 s; 0.83±0.08 m. Success; completion time; end-effector path | Frozen baseline: 100%; 25.60±2.31 s; 1.81±0.12 m. Reported time/path reductions are 63.8%/54.1%. Success is saturated; the finding concerns execution quality. Exact paired-test values are not supplied. e09e10 |
| BananaInBowlTask component ablation Same local protocol; all seven displayed EWAM conditions have 100% success; Table 6 supplies point estimates. | Full 9.27 s; no self-learning 46.53 s; no correction 18.20 s; no memory 10.60 s; no rollback 10.27 s. Completion time | Memory-only 12.67 s versus update-only 14.67 s. Experience accumulation has the largest removal effect. Memory-only beating update-only does not establish update harm or a general ordering across tasks; the conditions change available mechanisms. e09e11 |
| BananaInBowlTask execution faults 125 simulation evaluations per model; reported mean±95% CI. | EWAM 2.2±0.6. Total faults per episode | Baseline 13.5±2.1; reported reduction 83.7%. Preserved as reported: displayed category means sum to 1.7 and 11.0, respectively, not the total column. The missing accounting is unresolved. e12 |
| RoboLab transfer and stacking Same 5×25 protocol per task; three-task subset totals 375 evaluations per model; reported 95% CIs. | Bin transfer: 100%, 11.20±1.12 s. Stacking: 94.4%±2.1%, 28.60±3.10 s. Success and completion time | Baseline bin transfer: 100%, 18.40±1.85 s. Stacking: 87.2%±3.1%, 42.30±5.20 s. Stacking gains 7.2 percentage points of success. This extends evidence within the simulation subset, without isolating which recovery component causes the gain. e09e14 |
| Inference overhead Dual RTX 5880 Ada workstation, 48 GB/GPU; 100 inference trials; mean±SD. | EWAM 76.4±2.5 ms; 1820 MB. Latency and GPU memory | Baseline 66.2±2.4 ms; 1690 MB: +15.4% latency and +7.7% memory. Faster task completion coexists with slower inference. The path exceeds a 20 Hz controller's 50 ms budget; buffered or asynchronous control is required for that target. e13 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 6. Removing experience accumulation has the largest reported effect on completion time. Original paper, p. 17 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the full-model row as the reference, then compare each removal while keeping the task and success column in view. All displayed conditions still reach 100% success, so the informative columns are time and path. The no-self-learning row is described as preventing accumulation of filtered experience; it should not automatically be read as disabling only optimizer steps. The bottom two rows separate retrieval without parameter updates from parameter updates without retrieval. There are seven EWAM rows including the full model, despite prose referring to seven ablation variants. Table 6 gives point estimates, so this crop does not imply uncertainty estimates absent from the table. e09e11e17e20
What it supports. Full EWAM takes 9.27 seconds, compared with 46.53 without self-learning and 18.20 without correction. Memory-only takes 12.67 seconds, below update-only at 14.67. The pattern supports the utility of accumulated experience and correction in this setting, while the full method benefits from combining mechanisms.
Where the evidence stops. The surrounding figures are inconsistent: the p. 18 qualitative image has a quantitative caption, while the p. 19 ablation chart has a qualitative caption and says 20 episodes. The text specifies 125 trials; uncertainty labeling also conflicts. No filter-disabled row isolates filtering.
7. Analysis & limitations
7.1 What the evidence leaves open
No physical-robot or sim-to-real result is reported. Recovery depends on simulator checkpoints, interfaces are backbone-specific, and compact memory remains vulnerable to occlusion, novel dynamics, ambiguous objects and long-horizon errors. e15
Figures 4 and 5 have swapped content/caption types. The ablation graphic says 20 episodes, while the tables describe 125 trials; the Figure 4 caption calls uncertainty both CI and SE. Table 6 point estimates are retained without inventing error bars. e11e17
Figure 1's module/update order does not match the formal loop. Figure 3 depicts a smoothness lower bound and force interval, whereas Eq. (52) specifies an EE SPARC upper bound and maximum force. Their equivalence is not established. e08e19
The filter's necessity is asserted from its design rather than isolated by a filter-disabled ablation. Capability comparisons with other named models are not matched empirical rankings. e18e20
7.2 Questions for discussion
- Does the memory advantage arise on the first rollout or only after admitted experience accumulates? [e09, e11, e16]
- Can filtering suppress harmful updates without starving the harder stacking task of useful experience? [e08, e14, e20]
8. Reproducibility audit
8.1 Requirements and known gaps
Preserve seeds 42/123/456/789/1024, deterministic matched trial order and memory reset per seed. Obtain the exact checkpoint, task/asset/simulator versions, module implementation, metric scripts and per-seed logs before attempting numerical replication. e09e16
Resolve offline data and training duration, initial memory contents, adapter reset policy, update cadence/minibatch size, action horizon/controller timing, dynamics-prior implementation and aggregate anomaly-score definition. The supplied specification does not determine these completely. e02e04e05e06e13e16
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Separate retrieval benefit from accumulation and optimizer benefit
Reader-proposed check: run a 2×2 retrieval-on/off and adapter-update-on/off comparison on the same per-seed trial sequences, while keeping routing, correction, rollback and filtering identical. Reset and log the initial memory and adaptable weights for every condition; use a common admitted replay stream in a controlled pass, followed by an on-policy pass. Report first-trial performance and per-trial time/path curves, not just averages. A retrieval advantage under identical replay would support the memory mechanism; an advantage appearing only with different experience streams would implicate data accumulation. Compare paired seed-level summaries. This check has not been run. e03e06e09e11e16
Check 2: Test whether the admission gate prevents degradation or merely removes hard cases
Reader-proposed check: in RoboLab, replay an identical logged mixture of successful, inefficient and faulty trajectories into full-gate, safety-only-gate and unfiltered memory/update conditions. Hold the backbone, initial adapters, training budget and evaluation sequences fixed; additionally match admitted sample counts by subsampling to separate data quality from quantity. Evaluate on held-out layouts without further learning, measuring time, collisions, empty grasps and the fraction of experience rejected, including stacking. The filter's protective claim predicts fewer harmful updates at matched data volume; stalled learning or equivalent faults would weaken it. Resolve the Figure 3/Equation (52) discrepancy before choosing thresholds. This is a proposed experiment, not a reproduced result. e08e09e12e14e16e20
8.3 Reading coverage
Visual audit: Inspected the title/author/version block; all five numbered figures, including mismatched Figure 4/5 captions; Tables 1–11; formal method, filtering and training equations; evaluation statistics; latency/hardware details; and Appendices A–C. All six final original crops were viewed, including the corrected architecture and multi-task bounds. The illustrations preserve source markings. Figure 1 ordering, Figure 3 threshold definitions, the Figure 4/5 content swap, the chart's 20-episode note, and CI/SE ambiguity are disclosed. Pages 2 and 22–23 were read as supplied text and are not claimed as visually inspected. Separate supplemental availability remains unverified; code and experiments are outside this reading.
PDF pages inspected for this edition: 1, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 24, 25, 26. 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 and title metadata (p. 1)
- Sections 1–2.1: introduction, related work and comparison scope (pp. 1–3)
- Sections 3.1–3.6: architecture, memory, anomaly detection, routing and correction (pp. 3–8)
- Sections 3.7–3.11: online loop, objectives, rollback, filtering and memory (pp. 8–13)
- Sections 4.1–4.9: protocol, latency, results, ablations, faults and multi-task evaluation (pp. 14–20)
- Sections 5–5.1: discussion, limitations and conclusion (pp. 20–21)
- References (pp. 22–23)
- Appendix A: reproducibility (p. 24)
- Appendix B: rollout and update details (pp. 24–25)
- Appendix C: interpretation and residual failures (pp. 25–26)
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's figure limitation was addressed by inspecting original PDF pages and all six final crops. References were read as supplied text; cited works were not independently reviewed.
- Only the supplied arXiv:2606.12690v1 artifact, dated 10 June 2026, was reviewed. Its title and two authors agree with the catalog; no revision or alternate edition was supplied for comparison.
- Code, checkpoints, rollout logs and experiments were not inspected or reproduced. Source illustration/caption discrepancies and unspecified implementation details remain unresolved.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author block, arXiv margin and abstract
The title matches the catalog; credits are Xin Zhou (Astronex Robotics) and Cong Miao (Nanjing University of Information Science and Technology). The artifact is arXiv:2606.12690v1, 10 June 2026.
Go to primary source ↓e02PDF pp. 3–5, Sections 3.1–3.2, Eqs. (1)–(10)
Visual observation, proprioception and language enter VLE and AR reasoning, followed by a memory-augmented DiT with shared-hidden-state state/action heads. Backbone parameters are frozen; inserted modules and adapters can change. Zero-shot excludes additional target-task demonstrations.
Go to primary source ↓e03PDF p. 6, Section 3.3, Eqs. (11)–(14); p. 7, Table 2; p. 13, Section 3.11, Eqs. (53)–(56)
Memory attention is injected residually at DiT layer 6; retrieval uses five items and gate 0.4. The proposed memory contains keys, execution values, outcomes and rollback anchors, with task/scene/layout/success retrieval weights 0.35/0.30/0.20/0.15.
Go to primary source ↓e04PDF p. 7, Section 3.4, Eqs. (15)–(21); p. 11, Section 3.9, Eqs. (36)–(39), (44)–(47)
Decision-time diagnostics use a previous residual, candidate dynamics inconsistency and four sigmoid risk heads. The realized residual is available after execution. Hallucination uses grasp confidence above 0.85 with force below 0.15 N. The later gated-system equations list a realized-error trigger before the transition.
Go to primary source ↓e05PDF p. 8, Sections 3.5–3.6, Eqs. (22)–(26); p. 7, Table 2; p. 25, Appendix B, Table 10
A supervised softmax/argmax router selects direct, conservative or rollback behavior. Reported anomaly regions are below 0.3, [0.3,0.7), and at least 0.7. Conservative correction has strength 0.5 and maximum joint correction 0.3 rad. Rollback uses simulator checkpoints, with current-state replanning if no stable state exists.
Go to primary source ↓e06PDF pp. 10–11, Section 3.8, Eqs. (27)–(35), and Eq. (51); p. 7, Table 2; p. 24, Appendix A
Training combines action, state, memory, anomaly, routing and correction losses with weights 1.0/0.5/0.3/0.8/0.6/1.2. Supervision can come from simulator labels and recovery targets. Online learning uses only admitted replay; learning rate is 10^-4, adapter rank 16, and buffer size 1000.
Go to primary source ↓e07PDF pp. 8–10, Section 3.7 and Figure 2; p. 25, Appendix B
The textual loop generates candidates, diagnoses and routes them, corrects and executes actions, then filters experience. Memory changes immediately upon admission and parameter updates are intermittent. Figure 2 depicts the post-execution anomaly/recovery and rejection branches.
Go to primary source ↓e08PDF p. 12, Section 3.10, Eq. (52); p. 13, Figure 3
All six conditions are required: SafetyScore at least 0.75, TaskScore at least 0.80, maximum gripper force at most 0.80, EE SPARC at most -2.5, path at most 1.2 m and time at most 15 s. Figure 3 instead depicts a generic smoothness lower bound and a force interval; no mapping to Eq. (52) is supplied.
Go to primary source ↓e09PDF pp. 14–15, Sections 4.1, 4.4–4.5; p. 24, Appendix A and Table 9
Matched RoboLab conditions use position perturbation sigma 0.02 m, lighting variation ±15%, and five named seeds with 25 trials each per task. Memory resets each seed; trial sequences match across models. The source reports mean ±1.96 SE and proposes paired seed-level tests.
Go to primary source ↓e10PDF p. 16, Table 5 and Section 4.6; p. 14, Table 3
BananaInBowlTask: both models have 100% success; baseline/EWAM time is 25.60±2.31/9.27±0.94 s, path 1.81±0.12/0.83±0.08 m, collisions 3.2±0.8/0.3±0.2, empty grasps 1.4±0.3/0.1±0.1, rollback 0/1.2±0.4. Table 5 labels intervals as 95% CIs; exact test values require logs.
Go to primary source ↓e11PDF pp. 16–17, Section 4.7 and Table 6
All seven displayed EWAM conditions reach 100% success. Full, no-memory, no-self-learning, no-correction, no-rollback, memory-only and update-only times are 9.27, 10.60, 46.53, 18.20, 10.27, 12.67 and 14.67 s; respective paths are 0.83, 0.95, 4.63, 1.42, 0.89, 0.99 and 1.07 m. Table 6 gives no intervals.
Go to primary source ↓e12PDF p. 18, Section 4.8 and Table 7
Reported total faults are 13.5±2.1 for the baseline and 2.2±0.6 for EWAM, with 95% CI labels. The four displayed category means are 3.2/1.4/2.1/4.3 and 0.3/0.1/0.4/0.9, which do not exhaust the stated totals; the difference is unexplained.
Go to primary source ↓e13PDF p. 14, Section 4.3; p. 15, Table 4; p. 24, Table 9
Latency over 100 trials is 66.2±2.4 ms baseline and 76.4±2.5 ms EWAM (mean±SD), with GPU memory 1690/1820 MB and throughput 621/520 step/s. The workstation has two RTX 5880 Ada GPUs, 48 GB each. The text acknowledges that 76.4 ms exceeds a 50 ms servo budget.
Go to primary source ↓e14PDF p. 20, Table 8 and Section 4.9
Three tasks yield 375 evaluations per model. BananasInBinOneMoreTask retains 100% success and reduces time from 18.40±1.85 to 11.20±1.12 s. BlockStackingOrderAgnosticTask changes success from 87.2%±3.1% to 94.4%±2.1%, time 42.30±5.20 to 28.60±3.10 s, and path 3.85±0.31 to 2.41±0.22 m; intervals are labeled 95% CIs.
Go to primary source ↓e15PDF pp. 20–21, Sections 5–5.1; p. 26, Appendix C
All validation is simulation-only, with limited tasks, layouts and backbone scope. Real-robot rollback is unvalidated; the compact memory does not implement the full proposed multimodal signatures. Occlusion, changed dynamics, object ambiguity, long horizons and sensor noise remain failure modes.
Go to primary source ↓e16PDF pp. 24–25, Appendices A–B, Tables 9–10 and surrounding text
Exact replication requires module code, task configurations, simulator/asset versions, checkpoint identity, metric scripts and per-seed logs. The source specifies deterministic sequences and memory resets, but does not fully specify offline data, update cadence, initialization, or all controller and model details.
Go to primary source ↓e17PDF p. 18, Figure 4 image/caption; p. 19, Figure 5 image/caption; p. 17, Table 6
Figure 4 contains qualitative failure panels although its caption describes quantitative ablations and calls error bars both 95% CIs and standard errors. Figure 5 contains the ablation bar chart although its caption describes qualitative failures. Its graphic says 20 episodes and has no visible uncertainty bars, unlike the main 125-trial protocol.
Go to primary source ↓e18PDF p. 3, Section 2.1 and Table 1; p. 25, Appendix C
Other named VLA/WAM systems are compared at capability level, not under common measured protocols. The empirical control is the same frozen Cosmos3 policy interface without EWAM layers.
Go to primary source ↓e19PDF p. 6, Figure 1 and caption; pp. 4–5, Eqs. (1)–(10) and Section 3.2; p. 9, Section 3.7
Figure 1 is a schematic with strategy selection before mismatch detection and self-learning before memory writing. The equations and textual loop instead specify anomaly detection before routing and qualified memory admission before scheduled updates; memory is inserted inside the DiT.
Go to primary source ↓e20PDF p. 26, Table 11, filtering-necessity row; p. 17, Table 6; p. 12, Section 3.10
The source calls filtering necessary, citing its design role. Table 6 does not include a filter-disabled comparison, so the empirical necessity of the six particular thresholds is not isolated.
Go to primary source ↓8.5 Primary sources
EWAM: An Enhanced World Action Model for Closed-Loop Online Adaptation in Embodied Intelligence ↗
PDF · 8,046 extracted words
Source fingerprint
01e34604463f55790c28f03aac545ac1a97c3acd2502bc8f5ed59648de4c878b