AttenA+: Rectifying Action Inequality in Robotic Foundation Models
1. Paper overview
In one sentence: AttenA+ gives slow demonstrated actions greater relative training weight, improving several manipulation policies while depending on a task-specific link between speed and precision. e-probleme-velocitye-objectivese-liberoe-wam-traininge-physicale-evaluation
| At a glance | What to know |
|---|---|
| Research problem | Author claim Uniform timestep loss weighting can underemphasize precision-sensitive interaction phases. The authors propose demonstration speed as an inexpensive proxy for criticality; the assumed link between slowness and importance is the central hypothesis, not an independently measured information-theoretic quantity. e-problem |
| Core mechanism | Source description A parameter-free training-objective modification applies the same timestep weights to regression, flow matching and a formulated diffusion extension, without changing the host backbone. e-weighte-objectives |
| A key reported result | LIBERO four-suite manipulation with OpenVLA-OFT: 98.60; Spatial 99.0±0.16, Object 100±0.00, Goal 98.8±0.28, Long/10 96.6±0.30. Average task success rate (%). Separate suite models; best checkpoint selected on the test split; evaluation across four random seeds. OpenVLA-OFT 97.10 overall; Long/10 94.5. Gains are 1.50 and 2.1 percentage points respectively. Strong reported success near the benchmark ceiling; test-based checkpoint selection limits an unbiased generalization interpretation. No aggregate standard deviation is supplied. e-liberoe-evaluation |
| Reading caution | Source description The authors explicitly exclude a universal slow-equals-critical assumption: fast ballistic interactions can be decisive, and velocity omits force/torque information. e-limitations |
Core contributions
- Source description
A parameter-free training-objective modification applies the same timestep weights to regression, flow matching and a formulated diffusion extension, without changing the host backbone. e-weighte-objectives
- Source description
Evaluations cover OpenVLA-OFT and pi0.5 on LIBERO, a frozen-backbone Fast-WAM adaptation on RoboTwin, and physical Franka manipulation. They test policy success rather than video plausibility. e-liberoe-robotwine-wam-traininge-physical
Figure 3. Demonstration speed changes which action errors receive training priority. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left side as the host policy's visual-language-to-action mapping, then follow the right side from velocity v_t to weight w_t. Orange labels emphasize selected interaction frames; turquoise arrows identify lower-weight motions. The dashed expansion from the model's Actions box can suggest predicted-action weighting, but Section 3.3.1 and Eq. (2) explicitly compute speed from ground-truth demonstrations. Likewise, the plus sign beside Original Actions is schematic: Eq. (4) multiplies prediction error by w_t. It does not add weights to commands. This distinction is essential when translating the drawing into training code. e-velocitye-weighte-objectives
What it supports. The mechanism changes optimization priority while retaining the existing action predictor. A scalar weight derived from a demonstrated timestep is shared across that timestep's action dimensions. The intended benefit is greater learning pressure on slow alignment, grasping and placement phases without adding trainable attention parameters.
Where the evidence stops. The diagram omits clipping and optional normalization. Its schematic curves and 'attended actions' should not be interpreted as signed speed measurements or a test-time controller. Equations (2)–(4), rather than the ambiguous drawing alone, specify the implementation.
2. Motivation
2.1 The problem and the proposed response
Uniform timestep loss weighting can underemphasize precision-sensitive interaction phases. The authors propose demonstration speed as an inexpensive proxy for criticality; the assumed link between slowness and importance is the central hypothesis, not an independently measured information-theoretic quantity. e-problem
2.2 What this reading follows
A robot can traverse free space successfully yet fail during the final alignment or release. AttenA+ asks whether training should spend equal effort on those phases. Its answer is a small intervention in the loss: derive a scalar weight from demonstrated motion speed and apply it to each timestep's action error. The same idea is tested with regression, flow-based policies and a Fast-WAM adaptation. This reading follows the weight construction, then separates the reported success gains from the assumptions behind them. The key questions concern the velocity proxy, configuration sensitivity, checkpoint selection and deliberately slowed physical demonstrations. e-probleme-velocitye-objectivese-liberoe-wam-traininge-physicale-evaluation
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 | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is VLA. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded optimization/policy classification fits a loss-weighting framework. AttenA+ defines no new joint future/action architecture, inverse-dynamics pathway or control quadrant. Applying it to Fast-WAM does not make AttenA+ itself a One Model WAM, and the experiments do not measure inference acceleration. e-weighte-objectivese-wam-training
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Follow the expert target into the loss, not into a new controller
The easiest way to implement the idea conceptually is to keep two paths separate. Observations and language enter the existing policy to produce predicted actions. Expert actions provide both the supervised target and the speed used to construct a timestep weight. For LIBERO, the weight ignores binary gripper state when measuring motion, but Eq. (4) broadcasts it across every action dimension when scoring error. Thus even gripper prediction can receive extra emphasis during a slow phase. This follows from the equation rather than from an added gripper module. Figure 3 compresses these paths into an ambiguous drawing; its apparent addition of weights to actions should be read through the multiplicative loss formulation. At deployment, the source specifies no new weighting controller or future-imagination procedure. e-velocitye-objectives
Figure 8. Different mappings redistribute training emphasis in different ways. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the upper-left speed panel: horizontal position is timestep and vertical position is trajectory index. Compare the same populated coordinates in the four mapping panels, then use the bottom-right uniform panel as a reference. Inspect each color bar separately. The inverse-squared panel contains broad regions at its upper displayed weight, while exponential decay places most entries near its lower displayed weight. The latter still gives relatively more weight to slower actions; its displayed weights do not exceed one. White space is outside the plotted trajectory support and should not be read as a measured zero weight. e-diagnostice-weighte-configuration
What it supports. This visual makes the training intervention concrete: identical demonstrations can yield broad emphasis or sparse relative emphasis depending on the chosen mapping. The inverse-squared rule produces extensive saturation at clip 2, so its effective behavior depends on clipping as well as the raw inverse-square formula.
Where the evidence stops. These are transformed demonstration quantities, not measurements of causal importance or task success. The panel explicitly uses normalized speed, but the exact speed normalization and its relationship to optional mean-one loss normalization are not fully specified.
5.2 Treat clipping and normalization as part of the method
An inverse-speed rule has a practical hazard: a nearly stationary demonstration can dominate the weighted objective. The paper addresses this by bounding weights and optionally restoring their mean to approximately one. These operations affect different things. Clipping limits concentration; normalization controls overall loss scale. As reader analysis, a comparison that changes both timestep emphasis and average loss magnitude can mix the proposed mechanism with an effective optimization-scale change. Figure 8 helps expose this distinction because the mapping panels use different ranges, with exponential weights mostly below one. A reproduction therefore needs the precise preprocessing and normalization order, not just the four formulas. The source leaves zero-speed handling and complete per-checkpoint settings unclear, and Table 4 confirms that apparently reasonable mapping choices can reduce success. e-weighte-diagnostice-configuratione-ablation
5.3 Ask which form of transfer each experiment establishes
The experiments support several narrower conclusions. The OFT comparison tests the intervention on a strong regression policy, and the pi0.5 pair extends the evidence to a generative policy. RoboTwin tests adaptation of a pretrained Fast-WAM system through its action head while the backbone stays frozen; it does not establish improved learning of world dynamics. The Franka study provides physical execution evidence, with a tie on the simplest drawer task and gains elsewhere. However, those demonstrations intentionally encode precision demands through reduced speed. My reading is that the combined evidence supports a useful manipulation-training heuristic, not general criticality detection. Test-based checkpoint selection in LIBERO, absent uncertainty for other comparisons and the negative ablation results should all accompany the headline success rates. e-liberoe-wam-traininge-physicale-evaluatione-ablatione-limitations
5.4 Training and inference
During training
OFT trains separate suite models for 200,000 steps, saving every 5,000; each takes approximately 35 hours on one H800. Existing optimization and preprocessing settings are retained. e-oft-training
Fast-WAM freezes vision encoders and its pretrained backbone, updating only the final action head for one epoch: approximately four days on two H800 GPUs, with checkpoints every 2,000 steps. e-wam-training
During inference
The host policy still maps observations and instructions to actions: deterministic regression or generative action denoising. Ground-truth-derived weights affect training and introduce no specified test-time weighting, planning or world rollout. Controller timing and feedback cadence remain inherited and are not documented here. e-objectivese-velocitye-wam-training
5.5 Implementation flow
- Measure demonstrated motion
Compute a norm of expert action channels. For seven-dimensional LIBERO actions, the first six channels determine speed; binary gripper state is excluded. The source alternates between translational/rotational degrees of freedom and 'joint velocities', leaving physical units and channel scaling unclear. e-velocity
- Convert speed to optimization priority
Map speed through inverse, inverse-squared, exponential or reciprocal-log weighting. Clip extremes and optionally normalize to mean one. 'Attention' here means a loss multiplier, not learned token attention. e-weight
- Train the existing predictor
Broadcast each weight over the timestep's action dimensions. Multiply L1 regression error or the host generative squared-error objective, then optimize eligible parameters. Appendix B supplies a diffusion formulation but no separate diffusion experiment. e-objectives
6. Experiments & results
AttenA+ changes how robotic policies learn from demonstrations: slow action steps receive greater relative loss weight because the authors associate them with precision-sensitive manipulation. It adds no learned attention module. Reported simulation and Franka gains make the heuristic promising, but task-dependent ablations, checkpoint selection and deliberately speed-shaped demonstrations limit the generality of the evidence.
6.1 Read the original evidence
Table 1. Reported LIBERO success rises from 97.10% to 98.60%. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the last two rows because they compare the host model to its AttenA+ variant. Spatial, Object, Goal and 10 are separate suite columns; the average SR column summarizes success, while ER reports error. The 10 column corresponds to the paper's long-horizon evaluation. Read the plus/minus values only in the AttenA+ suite cells: Section 4.1 identifies them as variation across four evaluation seeds. The SR-I column on a baseline row expresses the improvement to AttenA+, not a property of that baseline alone. The wider leaderboard is useful context, but does not by itself establish identical training protocols. e-liberoe-evaluatione-reportinge-oft-training
What it supports. The average gain is 1.50 percentage points, with Long/10 increasing from 94.5% to 96.6%. All four displayed suite means improve against OpenVLA-OFT. These are task-success measurements, and the reported advantage is especially relevant where several manipulation stages must succeed in sequence.
Where the evidence stops. Appendix E.1 selects checkpoints on the test split, limiting generalization claims. The RER-R column prints negative reductions despite Eq. (6)'s positive-reduction definition; preserve the table and interpret its direct success/error columns first.
Table 2. RoboTwin gains are concentrated in clean scenes. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Clean and Rand. separately before interpreting the overall SR. They represent the fixed-scene and randomized conditions of the 50-task bimanual simulation benchmark described in Table 5. Compare AttenA+WAM directly with Fast-WAM: Appendix E.2 freezes the vision encoders and pretrained WAM backbone and fine-tunes only the final action head. The cross in Embodied PT. is the paper's embodied-pretraining category, not evidence that the entire system was trained from scratch. Other rows provide leaderboard context across systems with different histories. No uncertainty interval is attached to this table's success estimates. e-robotwine-reportinge-wam-training
What it supports. Clean success increases from 91.88% to 93.06%, whereas randomized success changes from 91.78% to 91.86%. Thus the aggregate result mainly reflects improvement in clean conditions. The paper reports 92.46% overall, but individual tasks can regress; randomized Handover Mic is one example in Table 5.
Where the evidence stops. Fast-WAM's displayed condition means average to 91.83%, whereas this table prints 91.80%. Its +0.6 gain is approximate, and the abstract uses 92.4 for AttenA+. Missing uncertainty prevents judging how stable the small randomized gain is.
Figure 6. Physical gains come from cube placement and the more complex tasks. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the left table to retain trial denominators: each task has 50 trials for each policy. Compare rows within a task before moving to the aggregate bars on the right. Close Draw is a tie; Put Cube, Multi-object and Long gain two, four and three successful trials respectively. The chart summarizes average success as 92.5 and 97, error as 7.5 and 3, and absolute improvement as 4.5 percentage points. Its last category is printed ERE-R and displays relative error reduction, so the value 60 is not another success rate. Table 6 supplies the individual binary outcomes. e-physicale-qualitative
What it supports. The physical results support a modest improvement on this four-task setup, with the largest absolute task gain in multi-object manipulation. They also demonstrate the distinction between percentage points and relative error reduction: raising average success by 4.5 points corresponds to reducing the reported error rate from 7.5% to 3%.
Where the evidence stops. Critical demonstrations were deliberately slowed to one-third speed. Robustness to arbitrary timing or unseen tasks is unestablished. The Figure 5 caption calls the drawer task opening; Section 4.4, the results label and Figure 13 identify closing.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO four-suite manipulation with OpenVLA-OFT Separate suite models; best checkpoint selected on the test split; evaluation across four random seeds. | 98.60; Spatial 99.0±0.16, Object 100±0.00, Goal 98.8±0.28, Long/10 96.6±0.30. Average task success rate (%) | OpenVLA-OFT 97.10 overall; Long/10 94.5. Gains are 1.50 and 2.1 percentage points respectively. Strong reported success near the benchmark ceiling; test-based checkpoint selection limits an unbiased generalization interpretation. No aggregate standard deviation is supplied. e-liberoe-evaluation |
| LIBERO four-suite manipulation with pi0.5 Generative backbone comparison in Table 3; a separate detailed pi0.5 training recipe is not supplied. | 97.95 Average task success rate (%) | pi0.5 96.85; +1.10 percentage points. Supports usefulness beyond direct regression. Per-seed uncertainty for this model pair is absent. e-liberoe-objectivese-oft-traininge-wam-training |
| RoboTwin 2.0 clean and randomized bimanual manipulation 50 simulation tasks; Fast-WAM action-head adaptation, with a frozen pretrained backbone and validation-selected checkpoint. | 92.46 overall; 93.06 clean and 91.86 randomized. Average task success rate (%) | Table 2 prints Fast-WAM 91.80 overall (91.88/91.78) and LingBot-VA 92.2; the paper labels gains +0.6 and +0.3 percentage points. Most gain is in clean scenes. Fast-WAM's two printed condition means average to 91.83, not 91.80; the abstract prints 92.4 for AttenA+. These precision inconsistencies remain unresolved. No seed uncertainty is reported. e-robotwine-reportinge-wam-training |
| Four physical Franka manipulation tasks Drawer closing, cube placement, multi-object and sequential manipulation; 50 trials per task per method. | 50/50, 50/50, 49/50, 45/50; average 97.0. Task successes / 50; average success rate (%) | OpenVLA-OFT 50/50, 48/50, 45/50, 42/50; average 92.5; +4.5 percentage points. Physical execution improves on three tasks and ties on drawer closing. Demonstration speed was deliberately reduced during critical phases; this is not a test of arbitrary human timing or a clearly documented held-out-task protocol. e-physicale-qualitative |
| LIBERO weighting strategy and clipping ablation OpenVLA-OFT; Table 4 varies mapping, with clips 2 across suites and 2/3/5 on Goal. | Long/10 at clip 2: exponential 96.8, inverse squared 94.2, log 88.8. Goal exponential: 99.0/95.4/97.4 at clips 2/3/5. Task success rate (%) | Baselines: Long/10 94.5; Goal 97.9. Weighting can help or hurt substantially. The default inverse-squared rule is not universally best; the source's prose about consistent improvement at moderate clips overstates this table. e-ablatione-weight |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. Weighting is useful only with an appropriate mapping and clipping choice. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down a suite column to compare mappings at the same clipping threshold. Read across the three Goal blocks to see how one mapping responds to clips 2, 3 and 5. Each block pairs success rate with its change from the italicized baseline. This arrangement matters: Spatial, Object and Long/10 are only shown at clip 2, so the table is not a complete threshold sweep for every suite. The red entries are informative counterexamples to universal improvement. On Long/10, compare exponential decay's positive result with inverse squared and logarithmic weighting, which underperform the baseline. e-ablatione-weighte-configuration
What it supports. Exponential weighting reaches 96.8% on Long/10 against 94.5%, while logarithmic weighting falls to 88.8%. On Goal, exponential weighting changes from 99.0% at clip 2 to 95.4% at clip 3. These contrasts show that neither a single mapping nor more permissive clipping guarantees improvement.
Where the evidence stops. Appendix C names inverse squared as the default, but this table does not identify the exact configurations behind every headline result. Seed uncertainty is absent. The authors' statement that moderate clips consistently improve performance is stronger than the displayed evidence.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors explicitly exclude a universal slow-equals-critical assumption: fast ballistic interactions can be decisive, and velocity omits force/torque information. e-limitations
Four LIBERO evaluation seeds do not remove selection bias from searching test-set checkpoints. RoboTwin aggregate gains also conceal regressions, such as randomized Handover Mic falling from 100 to 91. e-evaluatione-robotwin
Tables 1–2 print negative relative error reductions although Eq. (6) defines positive reductions when error decreases. Some comparator summaries differ from Table 5. Use direct success rates and preserve these discrepancies rather than silently correcting the published tables. e-reporting
7.2 Questions for discussion
- Does velocity alignment matter beyond changing gradient scale and emphasizing repeated near-static frames?
- How much of the reported gain persists with validation-only model selection and demonstrations that do not deliberately encode criticality in speed?
8. Reproducibility audit
8.1 Requirements and known gaps
Recover exact baseline optimization settings, action units/scaling, zero-speed handling, padding masks, normalization order and each checkpoint's mapping/clip choice. Appendix E mentions clips 2 and 5, while the ablation also uses 3. The flow equations omit a fully specified interpolation-time/noisy-input construction. e-configuratione-objectivese-ablatione-oft-training
Physical replication requires the documented 50/100 demonstration allocation, one-third-speed critical phases, waiting-frame removal and smoothing; their exact filtering/smoothing parameters are absent. Two H800 GPUs were used for training and an RTX-4090 for deployment. e-physical
Proposed checks: compare uniform, velocity-aligned and shuffled weights under identical mean loss scale and validation-only checkpoint selection; separately test whether benefits survive demonstrations whose timing no longer signals criticality. Neither experiment was run for this report. e-weighte-evaluatione-physicale-limitations
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Separate velocity alignment from loss-scale and checkpoint effects
Reader-proposed, not run: on LIBERO-10, train identical OFT models with uniform weights, exponential weights at clip 2, and within-trajectory shuffled versions of those same weights. Keep valid-timestep mean weight equal to one, optimization settings, data order, training budget and evaluation seeds matched. Select checkpoints on a distinct validation set and evaluate the test set only afterward. Report task success and held-out action error by speed range, with uncertainty across repeated runs. An advantage of aligned weights over both controls would support velocity-specific prioritization. Comparable performance from shuffled weights would weaken that explanation; loss of the gain after validation-only selection would expose protocol sensitivity. e-weighte-ablatione-evaluatione-oft-training
Check 2: Break the demonstration speed–criticality correlation
Reader-proposed, not run: collect matched demonstrations for the Franka multi-object task under two timing protocols: the paper's one-third-speed interaction phases and a safe protocol without deliberate interaction slowing. Match object configurations, demonstration counts and waiting-frame removal; document smoothing precisely. Within each dataset compare uniform and inverse-squared weighted training, with predeclared clipping and normalization, identical initialization and checkpoint selection. Evaluate on the same held-out scenes, recording task success and contact/alignment failure locations. A benefit confined to deliberately slowed data would indicate dependence on timing cues. A retained benefit without that cue would motivate investigating what other trajectory structure the weights capture. e-physicale-weighte-configuratione-limitations
8.3 Reading coverage
Visual audit: Visually inspected the title/version and author block; all method and objective pages; Tables 1–6; Figures 1–13; limitations; Appendix E hardware, frozen modules, schedules and checkpoint selection; and Appendix F detailed outcomes. Every final crop was inspected with its headers, axes and legends retained. Figure 3 was checked against Eqs. (2)–(4), and metric/table and drawer-label discrepancies are disclosed. All five supplied text chunks were read, including references; reference-only p. 11 was text-read but not rendered. Separate supplements, linked code and external baseline documentation were not supplied for review.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract
- 1 Introduction
- 2 Related Works (2.1–2.3)
- 3 Methodology (3.1–3.3.3)
- 4 Experiment (4.1–4.4)
- 5 Conclusion
- References
- Appendix A: Preliminary Concepts (A.1–A.3)
- Appendix B: Diffusion Policy (B.1–B.2)
- Appendix C: Weighting Strategies
- Appendix D: Speed and Weight Visualizations
- Appendix E: Model Training (E.1–E.2)
- Appendix F: Detailed Evaluation; Tables 5–6 and Figures 11–13
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.
- Revision scope: the supplied 22-page PDF is arXiv:2605.13548v3, dated 1 June 2026 and marked 'Preprint. Under review.' This bundle's title, all ten authors and six affiliations are verified from its title page (e-identity). Earlier versions were not supplied, so revision-to-revision scientific changes cannot be assessed.
- Text extraction did not reconstruct figure images; all thirteen figures and six tables were separately inspected in the supplied PDF. All five text chunks were read in full.
- Separate supplemental material availability has not been fully verified; no separate supplement was supplied.
- The linked code and external baseline documentation 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, author/affiliation block and arXiv margin
The title page reads 'AttenA+: Rectifying Action Inequality in Robotic Foundation Models' and lists Daojie Peng, Fulong Ma, Jiahang Cao, Qiang Zhang, Xupeng Xie, Jian Guo, Ping Luo, Andrew F. Luo, Boyu Zhou and Jun Ma. The artifact identifies itself as arXiv:2605.13548v3, 1 June 2026, preprint under review. Its six affiliations are HKUST(GZ), HKU, USTC, IDEA Research, SUSTech and X-Humaniod.
Go to primary source ↓e-problemPDF pp. 1–4, Introduction, Sections 3.1–3.2 and Figure 2
The authors motivate temporal loss weighting by associating slow manipulation phases with precision demands and fast phases with transitions. This is a proposed physical heuristic, illustrated with trajectories and task frames.
Go to primary source ↓e-velocityPDF pp. 4–5, Eq. (2), Figure 3 and Section 3.3.1
Weights derive from ground-truth action magnitudes. LIBERO uses its first six continuous action dimensions and excludes binary gripper state. Figure 3 schematically connects model actions to a velocity field, whereas the equations and prose explicitly specify expert targets.
Go to primary source ↓e-weightPDF p. 5, Sections 3.3.1–3.3.2; pp. 13–14, Appendix C, Eqs. (12)–(15)
Four mappings are inverse, inverse squared, exponential decay with alpha 5.0, and reciprocal log(1+speed). Inverse squared is called the default. Weights are clipped to [1/clip_max, clip_max]; mean-one normalization is optional.
Go to primary source ↓e-objectivesPDF p. 6, Section 3.3.3, Eqs. (4)–(5); pp. 12–13, Appendices A–B, Eqs. (7)–(11)
The scalar timestep weight multiplies each action-dimension L1 error, flow-field squared error, or diffusion denoising squared error. Inputs are visual observations and language; targets are demonstrated actions. The diffusion extension is formulated here without its own reported benchmark.
Go to primary source ↓e-evaluationPDF p. 6, Section 4.1; p. 17, Appendix E.1
LIBERO results use a best-performing checkpoint followed by evaluation across four random seeds. Appendix E.1 explicitly selects the checkpoint using success on the corresponding test split.
Go to primary source ↓e-liberoPDF p. 7, Table 1, OpenVLA-OFT and AttenA+OFT rows; Table 3, both model pairs
Table 1 reports average SR 97.10 to 98.60 and Long/10 SR 94.5 to 96.6. AttenA+ suite means and deviations are 99.0±0.16, 100±0.00, 98.8±0.28 and 96.6±0.30. Table 3 reports pi0.5 average 96.85 to 97.95.
Go to primary source ↓e-robotwinPDF p. 7, Table 2; p. 18, Table 5 caption, Average row and Handover Mic row
RoboTwin is a 50-task bimanual simulation benchmark with clean and randomized scenes. AttenA+ reports 93.06/91.86 and mean 92.46; Fast-WAM reports 91.88/91.78, with mean printed as 91.80 in Table 2. Handover Mic randomized success falls from 100 to 91.
Go to primary source ↓e-reportingPDF p. 1, abstract; p. 6, Eq. (6); p. 7, Tables 1–2; p. 18, Table 5 Average row
The abstract uses 92.4 for RoboTwin, while Table 2 uses 92.46. Table 2 prints a rounded +0.6 gain and negative RER-R entries, although Eq. (6) defines a positive reduction when error falls. Tables 2 and 5 also differ for some comparator averages, including Motus clean and X-VLA.
Go to primary source ↓e-ablationPDF p. 8, Table 4 and Section 4.3
At clip_max 2, exponential weighting gives LIBERO-10 SR 96.8, inverse squared 94.2, and log 88.8, versus baseline 94.5. On Goal, exponential gives 99.0/95.4/97.4 for clips 2/3/5 versus 97.9. No strategy dominates.
Go to primary source ↓e-physicalPDF pp. 8–9, Section 4.4 and Figure 6; p. 19, Table 6
Four Franka tasks receive 50 trials per method. Baseline successes are 50/48/45/42 and AttenA+ 50/50/49/45. Easy and harder tasks use 50 and 100 demonstration trajectories respectively. Critical demonstration phases use one-third baseline speed; waiting frames are removed and actions smoothed. Fine-tuning uses two H800 GPUs and deployment an RTX-4090.
Go to primary source ↓e-limitationsPDF pp. 9–10, Conclusion
Authors acknowledge that critical actions can be fast in dynamic tasks and that velocity alone ignores force and torque.
Go to primary source ↓e-diagnosticPDF pp. 14–16, Appendix D and Figures 7–10, especially Figure 8 on p. 15
Heatmaps compare normalized speed and four weighting rules with a uniform baseline. Figure 8 uses clip_max 2 and separate color scales; its exponential panel ranges approximately 0.5–1, while inverse-based panels reach 2. The plots are weight diagnostics rather than controlled success measurements.
Go to primary source ↓e-oft-trainingPDF p. 17, Appendix E.1
Separate LIBERO suite models train for 200,000 steps, save every 5,000, and take about 35 hours each on one H800. Optimizer, learning rate schedule, batch size and preprocessing are inherited from OpenVLA-OFT without their numerical settings being restated.
Go to primary source ↓e-wam-trainingPDF p. 17, Appendix E.2
Vision encoders and the pretrained WAM backbone are frozen; only the final action head is fine-tuned for one epoch. Checkpoints are saved every 2,000 steps and selected by validation success. The reported duration is about four days on two H800 GPUs.
Go to primary source ↓e-configurationPDF p. 5, Section 3.3.2; pp. 13–15, Appendices C–D; p. 17, Appendix E opening and E.1–E.2
The source gives optional weight normalization, singular inverse/log formulas without an explicit zero-speed epsilon, normalized-speed diagnostic plots, a default inverse-squared rule, and general clip settings 2 and 5. It does not map complete preprocessing, normalization and clipping configurations to each headline checkpoint; Table 4 on p. 8 also tests clip 3.
Go to primary source ↓e-qualitativePDF p. 8, Figure 4; p. 9, Figure 5 caption; pp. 20–22, Figures 11–13
Selected frames label grasping/alignment stages as critical in LIBERO, RoboTwin and physical tasks. Figure 5's caption says drawer opening, while Section 4.4 and Figure 13 describe closing the open drawer. These examples are qualitative illustrations.
Go to primary source ↓8.5 Primary sources
AttenA+: Rectifying Action Inequality in Robotic Foundation Models ↗
PDF · 8,726 extracted words
Source fingerprint
45b5c6ccd976e20161f77ede0e3963bf73f3056c552b50d33b0ab563be281177