Bridge-WA: Predicting Where and How the World Changes for Robotic Action
1. Paper overview
In one sentence: Bridge-WA trades a deployment-time future generator for compact predicted outcome, change and motion priors, whose usefulness depends on how the action transformer reads them. e02e03e05e08e10e12e13e14e17
| At a glance | What to know |
|---|---|
| Research problem | Author claim Reactive vision-language-action policies may exploit appearance correlations without representing manipulation consequences. The authors seek useful foresight without deploying a large future-image generator. Their bottleneck retains what outcome is intended, where change should occur and how affected regions should move. e02 |
| Core mechanism | |
| A key reported result | Dobot Nova2 manipulation under visual shifts: Reported Easy 73.6%; Hard 69.1%. Normalized task score, labeled Avg. SR; includes partial credit. Five tasks, 50 demonstrations each; ten Easy trials and five trials per distractor, lighting and tablecloth factor. X-VLA: 69.6% and 58.0% (+4.0 and +11.1 percentage points). These are not all-or-nothing completion frequencies. PickGrape improves strongly, but StackBowls remains weak and drawer Hard ties X-VLA. e09e13e16 |
| Reading caution | Author claim Teacher coverage, short-horizon image-space priors and offline-cache construction cost limit the method. Geometry-aware and longer-horizon extensions remain future work. e20 |
Core contributions
Figure 1. A teacher supplies training targets; compact predictor heads supply the deployed policy. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the orange branch: the snowflake marks the teacher as frozen during cache construction and policy distillation, after its own training. The cache supervises three predictor heads. Follow the lower path from multi-view images and language through Florence-2 to the predicted future, change and flow representations. Their arrow enters the action transformer, alongside the encoded policy context. Proprioception and action/time tokens have a separate drawn input path. Finally, the action decoder produces a chunk. Removing the orange branch preserves the encoder, predictor heads, WorldBridge layers and decoder; it does not leave the heads acting alone. e03e04e07e08e17e20
What it supports. The figure locates the computational saving precisely: a large teacher and its future decoding are moved into preparation, while compact priors remain active during action generation. These priors are more than auxiliary training targets because their predictions feed the deployed transformer. The diagram itself supplies no measured speedup.
Where the evidence stops. The deployment label “Predictor heads only” abbreviates the world-prior component. The full policy remains. Figure 1 also omits an explicit proprioception-to-predictor arrow, although Section 3.2 includes robot state in the predictor context; that wiring detail is unresolved.
2. Motivation
2.1 The problem and the proposed response
Reactive vision-language-action policies may exploit appearance correlations without representing manipulation consequences. The authors seek useful foresight without deploying a large future-image generator. Their bottleneck retains what outcome is intended, where change should occur and how affected regions should move. e02
2.2 What this reading follows
A robot opening a drawer needs to anticipate an outcome, locate the changing region and move in a useful direction. Bridge-WA represents those requirements separately. It first adapts a large future teacher to manipulation, compresses the teacher’s predictions into three kinds of supervision, and trains lightweight heads alongside an action policy. Deployment retains the heads and conditioned policy. The experiments support stronger average manipulation scores, especially under several appearance shifts, while exposing weaknesses in camera-view and cross-category transfer. Read the figures as an argument about a predictive interface: the decisive question is whether its content and routing improve executed actions. e02e03e05e08e10e12e13e14e17
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 | Dual-system |
| Prediction paradigm | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Insufficient evidence to decide
WAM membership is supported by predicted world priors used during action inference. Separate teacher and policy support a dual-system description during training; deployment shares an encoder with lightweight prediction heads. IDM is less secure: the paper formulates imitation conditioned on predicted summaries, without an explicit inverse-dynamics objective over observed state transitions. Joint predictor/policy training alone does not establish a unified future-and-action generator. Preserve Q4 as the catalog snapshot with this unresolved interpretation. e03e05e07e08
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 1. Construct supervision that survives removal of the teacher
The teacher is a preparation tool with its own training history. Appendix C.1 first adapts Wan2.2-TI2V-5B on BridgeData V2 and then fine-tunes its DiT on the downstream dataset, keeping the VAE and text encoder frozen during that target adaptation. Once frozen, it predicts a future latent and decoded future view. Equation (8) compresses these outputs differently: pooling yields outcome tokens; pooled absolute latent differences yield change support; flow between current and predicted future images yields motion. The cache also stores alignment metadata, so a policy sample receives the intended camera/time/task targets. Lightweight heads then learn these targets jointly with action imitation. This sequence explains how future supervision can remain useful after removing the expensive generator, without treating teacher predictions as observed physical truth. e03e04e05e06e07e17
5.2 2. Let the action stream decide when to read each prior
WorldBridge keeps queries in the policy stream while expanding what those queries can read. Active priors contribute key/value memories; change and flow additionally bias attention logits. Equation (3) activates a modality only between its start layer and the exclusive end determined by its layer budget and transformer depth. Inactive priors are omitted, rather than projected everywhere and ignored afterward. The intended order moves from global outcome information toward spatial support and local displacement near decoding. Figure 2 also shows modulation from prior summaries. During training, Appendix B.4 specifies blended teacher/predicted guidance; deployment uses the predictor alone. The ablation therefore matters: receiving all three priors and receiving them through the best conditioning design are empirically different configurations, not interchangeable descriptions of one model. e07e08e14
5.3 3. Separate predictive plausibility, robustness and completed actions
Reader analysis: the strongest case combines complementary evidence instead of asking one visualization to prove everything. Teacher-target maps illustrate the intended bottleneck, but cannot establish the deployed predictor’s accuracy or causal selectivity. Table 5 tests whether prior content and conditioning improve manipulation under the ablation protocol. Table 3 tests robustness across specified perturbation dimensions and exposes a camera-view weakness. Table 4 adds physical execution, but its fractional scores reward intermediate completion on several tasks. RoboTwin extends evaluation to bimanual control using an Easy-trained checkpoint on Hard, and the appendix adds Franka Easy results. These experiments support useful future conditioning across several settings. They do not establish a universal causal representation, strict binary real-robot success at the reported aggregate rate, or a measured real-time advantage. e12e13e14e16e18e19e20e21
5.4 Training and inference
During training
Jointly optimize action imitation and prior distillation. Future tokens use squared-L2 regression; change and flow use L1; every branch also has cosine alignment. All six weights default to 0.1. Policy training reads the cache without teacher forward passes. e06e07
Appendix B.4 specifies a 0.5 blend of teacher and predicted guidance during training, modulation scale 0.1 and change/flow bias scales 1.0. This qualifies the main-text description of predicted-prior conditioning. Several action-head families are permitted without a complete head-specific recipe. e08e17
During inference
Encode current observations and instruction, predict priors, and generate actions through WorldBridge and the action decoder. Teacher, cache and future-image decoder are absent. Proprioception belongs to the formal context although Figure 1 draws its connection only to the action-token encoder. e03e07e17
This is future-conditioned action prediction rather than explicit candidate-action search in a deployed simulator. Chunk execution length before re-observation and the precise online denoising schedule are unspecified. e03e08e17
5.5 Implementation flow
- Acquire a robot-specific future teacher
Initialize Wan2.2-TI2V-5B, pretrain on BridgeData V2, then adapt to the downstream dataset. During target adaptation only the DiT changes; VAE and text encoder remain frozen. Freeze the resulting teacher before caching. e04e17
- Compress and predict the future
Pool predicted future latents into outcome tokens. Pool and normalize absolute current/future latent differences into change maps; estimate image-space flow to the decoded predicted future. Cache these targets with sample alignment metadata. Lightweight heads share encoded visual-language context and learn to recover them. e03e05e06e07
- Condition action generation progressively
Policy queries read ordinary self-attention memories plus active future/change/flow memories. Change and flow add attention-logit biases. Route outcome information earlier, spatial support in middle-to-late layers, and local motion near decoding; omit inactive priors. Figure 2 additionally shows FiLM-style modulation. e08
6. Experiments & results
Bridge-WA compresses a manipulation-trained future teacher into outcome tokens, spatial change maps and motion-flow maps. A policy predicts these priors from current context and reads them while generating action chunks. Reported gains concern executed manipulation, with improved averages but uneven transfer across perturbations and tasks. Efficiency is architectural; deployment latency is not measured (e03, e10, e12, e13, e20).
6.1 Read the original evidence
Table 1. Average success improves while cross-category transfer remains a weakness. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row across the five public evaluation categories before inspecting Avg. SR. The categories distinguish ordinary in-distribution execution from changes in object category, common-sense demands, instruction semantics and texture. Keep the two π0-Fast action representations separate: relative chunk and delta chunk are different reported configurations. The delta-chunk row is the strongest average baseline and the appropriate comparator for the headline SR gain. Bridge-WA’s strongest entries are not evidence of uniform dominance; the Cross-Cat. column visibly favors another method. Table 2, inspected separately, reports intention and progress metrics rather than another version of this success measure. e09e10e11e20
What it supports. Bridge-WA reports 52.8% average SR against 43.1% for π0-Fast delta chunk, a 9.7-percentage-point difference. Its 23.0% cross-category result is below that baseline’s 34.2%, so the supported conclusion is better average performance across the selected categories, with a substantial remaining transfer gap.
Where the evidence stops. These rows compare complete systems. They do not isolate WorldBridge, establish equal compute, or provide uncertainty intervals. The separate Table 2 additionally imports several baselines from a specified frozen-backbone setting; do not merge its protocol into this SR comparison.
Table 3. Strong appearance robustness coexists with a large camera-view deficit. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat columns as distinct evaluation perturbations. Camera changes viewpoint, Robot changes initial state, and the remaining columns vary language or scene appearance/layout. The source says none of the methods trains on LIBERO-Plus data; this is robustness evaluation with the base manipulation tasks retained. Read Bridge-WA’s row as a profile before comparing its reported average with OpenVLA-OFT. Lighting and background are favorable cases for a change-focused representation. Then move back to Camera: its strongest baseline differs from the average leader. This prevents a favorable aggregate from becoming an unsupported claim that all visual shifts are handled. e09e12e17e20
What it supports. Bridge-WA leads the reported average at 72.1%, compared with OpenVLA-OFT’s 69.6%. It reaches 96.7% on lighting and 94.3% on background, yet only 25.0% on camera perturbations, where π0-Fast reaches 65.1%. The result supports selective robustness and motivates attention to viewpoint-dependent geometry.
Where the evidence stops. Zero-shot here means no training on LIBERO-Plus, not absence of robot or base-task training. The table does not by itself isolate the representation’s causal effect or establish a common training budget across all methods.
Table 4, including scoring footnote. Real-robot gains are measured with a mixture of binary and partial-credit task scores. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the retained scoring footnote. A fractional numerator such as 13.3 is accumulated credit, not a fractional number of fully successful trials. PickGrape and drawer manipulation use binary scores; buttons, fruit collection and bowl stacking allow partial completion. Easy uses ten trials per task. Each of the three Hard factors uses five, and Hard pools their scores over fifteen trials. The lower-right block averages across tasks. Compare Bridge-WA against X-VLA within the same task and condition, then inspect the overall block. Appendix A states that initial states, instructions and reset procedures are matched across methods. e09e13e16e20
What it supports. Bridge-WA’s normalized Easy/Hard averages are 73.6%/69.1%, versus X-VLA’s 69.6%/58.0%. The Hard improvement is 11.1 percentage points. PickGrape rises from 10/15 to 14/15, but drawer Hard stays at 7/15 for both methods and StackBowls remains difficult; the average does not imply gains everywhere.
Where the evidence stops. The trial counts are small and uncertainty is unreported. These averages cannot be read as binary completion frequencies. Main-text and appendix Dobot image resolutions differ, so the precise preprocessing setup needs clarification for reproduction.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| VLABench five-category manipulation Ten tasks; front/secondary/wrist RGB, robot state and language; five public categories. | Reported SR 52.8%; IS/PS 71.2%/64.0%. Average SR; intention score (IS); progress score (PS) | SR: π0-Fast delta chunk 43.1% (+9.7 percentage points). IS/PS: X-VLA 70.2%/51.2%. Cross-category SR is 23.0% versus 34.2% for π0-Fast delta. Table 2 imports some baselines from a separate frozen-backbone training setting. e09e10e11 |
| LIBERO-Plus zero-shot robustness Seven independently perturbed evaluation dimensions; no LIBERO-Plus training. | Reported 72.1%. Average success rate | OpenVLA-OFT 69.6% (+2.5 percentage points). Robot-state, lighting and background scores are 92.8%, 96.7% and 94.3%; camera-view success is 25.0%, versus π0-Fast’s 65.1%. Robustness is dimension-specific. e09e12 |
| RoboTwin 2.0 bimanual manipulation 15-task subset; BRIDGE-WA trains on Easy only and directly evaluates the same checkpoint on Hard. | Reported Easy 79.67%; Hard 37.73%. Average success rate by split | X-VLA: Easy 68.53%; Hard 36.00%. The Easy advantage is much larger than the Hard advantage; domain randomization remains difficult. e18 |
| Dobot Nova2 manipulation under visual shifts Five tasks, 50 demonstrations each; ten Easy trials and five trials per distractor, lighting and tablecloth factor. | Reported Easy 73.6%; Hard 69.1%. Normalized task score, labeled Avg. SR; includes partial credit | X-VLA: 69.6% and 58.0% (+4.0 and +11.1 percentage points). These are not all-or-nothing completion frequencies. PickGrape improves strongly, but StackBowls remains weak and drawer Hard ties X-VLA. e09e13e16 |
| VLABench-5 world-prior and routing ablation Same five tasks and optimization budget; stated 100 episodes per task; four reported tracks. | Layered conditioning 50.4%. Reported Avg. SR | No-world 46.3%; all-three token combination 46.6%; attention-only 44.8%; no-gate 49.3%. Routing matters: adding all three alone barely improves the baseline. The source does not fully explain Avg. aggregation from the track columns. e14 |
| Franka FR3 Easy manipulation Five tasks on the additional robot platform; Easy evaluation only. | 73.0%. Reported average success rate | VLA-JEPA 49.0%; π0.5 47.0%. Supports a second-platform result, not Franka OOD robustness. Table-specific trial denominators remain unclear relative to the generic protocol. e16e19 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 5. Prior content and the mechanism that injects it must be evaluated separately. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read F, C and O as future, change and motion-flow priors. The upper block changes the included sources; the lower block keeps all three present while changing conditioning. The first row anchors the no-world comparison. Tracks 1–4 are in-distribution, cross-category, common-sense and semantic-instruction evaluation. Section 4.2 states that configurations use the same five tasks and optimization budget, with one hundred evaluation episodes per task. First compare flow-only with change-only; then compare the all-three row with layered conditioning. This sequence separates the question of useful information from the question of access to that information. e08e14e20
What it supports. The reported averages are 46.3% without world tokens, 46.6% with all three in the token-combination block, and 50.4% with layered conditioning. Attention-only reaches 44.8%. Thus merely providing additional priors is insufficient; the strongest configuration depends on the conditioning design. The layered/no-world difference is 4.1 percentage points.
Where the evidence stops. Layered conditioning is not an isolated test of layer order alone. Exact routing indices and gate implementation are missing. Reported Avg. values are preserved: the source does not fully explain their aggregation from the four displayed tracks or report uncertainty.
Figure 4(IV). Each prior benefits from a different spatial budget. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend twice: color identifies the prior, while circles/solid lines denote success rate and squares/dashed lines denote progress score. Small, Mid and Large are not shared resolutions. Future uses 1×1, 3×3 and 4×4; change uses 2×2, 8×8 and 16×16; flow uses 2×2, 8×8 and 16×16. Read each color within its own sweep. The broken vertical axis separates the lower SR band from the upper PS band, so the apparent vertical distance between those metrics should not be interpreted continuously. These are single-factor sweeps, rather than a joint search over all three grid sizes. e14e15
What it supports. Future SR decreases from 46.8% at 1×1 to 43.4% at 4×4. Change peaks at 8×8 with 49.4% SR and 62.8% PS. Flow SR rises from 45.2% at 2×2 to 49.0% at 16×16. The pattern favors compact outcome summaries and more spatial detail for motion.
Where the evidence stops. The sweeps motivate asymmetric allocation but do not prove a joint optimum or a monotonic benefit beyond the tested sizes. They also do not establish that changed spatial resolution affects only useful information rather than optimization or conditioning capacity.
7. Analysis & limitations
7.1 What the evidence leaves open
Teacher coverage, short-horizon image-space priors and offline-cache construction cost limit the method. Geometry-aware and longer-horizon extensions remain future work. e20
Teacher-target heatmaps and selected rollouts illustrate representations, without establishing that the deployed predictor isolates causal effects. Comparisons lack uncertainty intervals; no measured latency supports a numerical efficiency claim. e21e20
Main-text Dobot RGB is 640×480, whereas Appendix A says 480×480. Table 8’s 25% binary drawer baselines are not reconciled with the generic ten-trial protocol. Neither discrepancy is silently repaired. e09e16e19
7.2 Questions for discussion
- Does correct spatial alignment of predicted priors matter more than extra supervision and parameters?
- Can predicted-only training retain the benefit of teacher/predicted guidance blending?
- Would geometry-aware priors improve camera robustness without losing the deployment advantage?
8. Reproducibility audit
8.1 Requirements and known gaps
Reference caching uses four GPUs, batch 128, ten teacher diffusion steps and horizon 30; teacher inputs are 256×256 with VLABench future_index 30. Reproduction needs aligned camera/time/task keys, teacher checkpoints and downstream demonstrations. GPU models, optimizer/schedule, policy-backbone freezing, exact routing indices and the chosen action-head implementation remain incompletely specified. e04e06e08e17
The single-factor sweep motivates compact future tokens, intermediate change grids and finer flow grids, without establishing their joint optimum. Exact flow estimation, normalization/pooling, selected five-task ablation identities and execution/re-observation timing require clarification before exact replication. e05e08e14e15e17
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: Control prior alignment while isolating the routing schedule
Reader-proposed check, not performed: reconstruct the VLABench-5 setup once task identities and routing indices are clarified. Use the same demonstrations, teacher cache, action head, parameter budget, seeds and episode initializations. Compare no priors, correctly aligned priors with the stated coarse-to-fine order, and the same priors with the modality order reversed but equal layer budgets. Add a within-task sample-shuffled prior control. Report SR/PS per track with paired uncertainty. If shuffled or reversed priors match correctly aligned layered conditioning, the claim that structured future content and its intended routing drive the gain would be weakened. e05e06e08e14e15
Check 2: Test whether teacher-guidance blending creates a deployment gap
Reader-proposed check, not performed: hold cache, policy architecture, task data and optimization budget fixed while training the default 0.5 teacher/predicted guidance blend and a predicted-guidance-only condition. Evaluate both with predicted priors only under matched Dobot Easy and Hard initializations. Record prior error against cached targets, the published partial-credit score and separately annotated full-completion frequency. A third evaluation using cached teacher guidance on held-out observations may serve as an explicitly offline diagnostic, never a deployable result. A large oracle-to-predicted drop or an advantage for predicted-only training would expose a train/deployment conditioning mismatch. e05e07e08e13e16e17
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page, all method and experiment pages, and Appendix A–D pages, including Figures 1–8 and Tables 1–8. All six final original crops were opened and checked; the architecture crop was adjusted to retain its complete lower border, and Figure 4(IV) was rendered at 600 DPI to read its narrow legend, labels and axis break. The Table 4 crop retains its scoring footnote because it changes the interpretation of the numbers. Figure 1’s frozen teacher and training/deployment arrows were cross-checked against Sections 3.2–3.3 and Appendix B–C; the schematic proprioception omission and abbreviated deployment label are disclosed. The size-sweep values agree with Section 4.2. All supporting method, numeric, training, protocol and proposed-check pages are included above. Reference-only pages 10–14 were read in the complete text but not visually inspected. No separate supplement, code or external project material was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 15, 16, 17, 18, 19, 20, 21. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract
- 1 Introduction
- 2 Related Work
- 3 BRIDGE-WA Model, including 3.1–3.3
- 4 Experiments, including 4.1–4.3
- 5 Conclusion and Limitations
- References
- Appendix A: setup, A.1 task definitions, A.2 evaluation protocol
- Appendix B.1–B.4: teacher, cache, predictor, conditioning
- Appendix C.1–C.3: training, baselines, additional results
- Appendix D: additional visualizations
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.
- Identity/version: the title page identifies arXiv:2607.02195v1, 2 July 2026. Title and all six authors match the catalog; small-capital styling is not a title change. No other revision or separate edition was supplied or compared (e01).
- All six text chunks were read individually, covering all 21 pages. Text extraction does not reconstruct figure images; the retained PDF was visually inspected separately on pages 1–9 and 15–21. Reference-only pages 10–14 were read as text.
- Separate supplemental material availability has not been fully verified; none was supplied.
- Linked code and project resources were not opened. No code was inspected and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author/affiliation block and arXiv margin
The title matches the supplied observed title, with BRIDGE-WA typeset in small capitals. Authors are Yongjie Bai, Hanting Wang, Mingtong Dai, Qijun Zhong, Yang Liu and Liang Lin. Four affiliations are listed. The margin identifies arXiv:2607.02195v1, 2 Jul 2026.
Go to primary source ↓e02PDF pp. 1–2, Abstract and Section 1
The authors motivate compact outcome, change and motion representations to retain predictive information while avoiding deployment-time dense future generation.
Go to primary source ↓e03PDF p. 3, Figure 1 and Section 3; p. 4, Eq. (2) and Section 3.2
The Florence-2 visual-language backbone feeds lightweight prior heads and an action transformer. A frozen teacher/cache supplies training targets; deployment uses predicted priors. The text includes proprioception in the predictor context, although Figure 1 explicitly draws it only into the action-token encoder.
Go to primary source ↓e04PDF p. 4, Figure 2(I), Section 3.2; p. 16, Appendix B.1, Eq. (7)
The Wan2.2-TI2V-5B teacher predicts future observations from images, robot state and language. Reference inputs are 256×256 and VLABench future_index is 30.
Go to primary source ↓e05PDF p. 16, Appendix B.1, Eq. (8) and definitions
Future targets pool predicted future latents; change targets normalize pooled absolute current/future latent differences; motion targets resize flow estimated between current and teacher-predicted future images. Exact pooling, normalization and flow implementations are not supplied.
Go to primary source ↓e06PDF p. 17, Appendix B.2, Eq. (9)
Cache entries contain sample key, future offset, priors and alignment metadata including camera order, time and task. The reference uses four GPUs, batch 128, ten teacher diffusion steps and a 30-action horizon, with half precision where possible. GPU models and total cache cost are unspecified.
Go to primary source ↓e07PDF p. 17, Appendix B.3, Eqs. (10)–(12)
Predictor heads share encoded visual-language context and train jointly with action imitation. Future regression uses squared L2; change/flow use L1. All branches also use cosine distance; each of six default weights is 0.1. Inference uses only predicted priors.
Go to primary source ↓e08PDF p. 4, Figure 2(II); p. 5, Section 3.3, Eqs. (3)–(4); pp. 17–18, Appendix B.4, Eqs. (13)–(15)
Policy queries read self and active prior key/value memories, with additive change/flow biases. Future enters earlier, change in middle-to-late layers and flow near decoding; inactive priors are omitted. Defaults: modulation 0.1, both bias scales 1.0, teacher/predicted guidance blend 0.5 during training. Numerical layer ranges are unspecified.
Go to primary source ↓e09PDF p. 5, Section 4, simulation and real-robot setup
VLABench uses front, secondary and wrist views and five categories. LIBERO-Plus perturbs seven dimensions without training on its data. Dobot has five tasks, 50 demonstrations each, ten Easy trials and five per Hard factor. Main-text RGB is standardized to 640×480.
Go to primary source ↓e10PDF p. 7, Table 1, all rows and Avg. SR/Cross-Cat. columns; p. 19, Table 7
VLABench average SR is 52.8 versus 43.1 for π0-Fast delta chunk. BRIDGE-WA category SRs are 78.0, 23.0, 51.1, 67.0 and 45.0; the strongest cross-category entry is 34.2. Table 7 reports ten individual tasks.
Go to primary source ↓e11PDF p. 7, Table 2, Avg. columns and caption
Average IS/PS are 71.2/64.0 for BRIDGE-WA and 70.2/51.2 for X-VLA. Diamond-marked baselines come from ACoT-VLA’s 60K-step frozen-backbone setting; matching categories/metrics does not establish identical training budgets.
Go to primary source ↓e12PDF p. 8, Table 3, BRIDGE-WA, OpenVLA-OFT and π0-Fast rows and caption
LIBERO-Plus average SR is 72.1 for BRIDGE-WA versus 69.6 for OpenVLA-OFT. BRIDGE-WA scores 25.0 on camera, 92.8 on robot state, 96.7 on lighting and 94.3 on background; π0-Fast scores 65.1 on camera.
Go to primary source ↓e13PDF p. 8, Table 4, per-task entries, Avg. SR block and scoring footnote
Dobot Easy/Hard averages are 73.6/69.1 for BRIDGE-WA versus X-VLA’s 69.6/58.0. PickGrape Hard is 14/15 versus 10/15; StackBowls Hard is 5/15 versus 5.1/15; drawer Hard is 7/15 for both. Several tasks use partial credit.
Go to primary source ↓e14PDF pp. 7–8, Section 4.2; p. 9, Table 5 and caption
VLABench-5 uses the same five tasks/optimization budget and states 100 episodes per task. Reported Avg. SR: no-world 46.3, change-only 48.4, flow-only 45.2, all-three 46.6, attention-only 44.8, no-gate 49.3, layered 50.4. Tracks 1–4 denote in-distribution, cross-category, common-sense and semantic-instruction. Avg. aggregation is not fully explained.
Go to primary source ↓e15PDF p. 6, Figure 4(IV), complete plot and caption; p. 8, Section 4.2, World-prior Size Sweep Analysis
Future 1×1 has SR/PS 46.8/60.2 versus 43.4/56.4 at 4×4. Change peaks at 8×8 with 49.4/62.8; flow SR rises from 45.2 at 2×2 to 49.0 at 16×16. The plot uses modality-specific sizes and a broken vertical axis.
Go to primary source ↓e16PDF pp. 15–16, Appendix A overview, A.1 and A.2
Dobot Nova2/Franka FR3 data use seven-dimensional Cartesian pose/gripper vectors, 30 Hz demonstrations and 50 demonstrations per task. Dobot D455/D405 videos are 480×480, differing from p. 5. Easy has ten trials; each Hard factor has five. Buttons allow 0/0.5/1, fruits/bowls 0/0.3/0.6/1; other tasks are binary. Initial states, instructions and resets are matched.
Go to primary source ↓e17PDF p. 18, Appendix C.1, Training Recipe
The teacher is pretrained on BridgeData V2, then fine-tuned on the downstream dataset before caching. Target fine-tuning updates only the 5B DiT; VAE, text encoder and other components stay frozen. Policy imitation/distillation follows with action horizon 30. Optimizer, learning-rate schedule, policy training steps and exact action-head implementation are unspecified.
Go to primary source ↓e18PDF p. 6, Figure 4(III); p. 18, Appendix C.3; p. 19, Table 6
RoboTwin 2.0 covers 15 tasks; BRIDGE-WA trains only on Easy and directly evaluates the same checkpoint on Hard. Average SR is 79.67/37.73, versus X-VLA’s 68.53/36.00 in Figure 4(III).
Go to primary source ↓e19PDF p. 19, paragraph preceding Table 8 and Table 8
Franka Easy average is 73.0% for BRIDGE-WA, 49.0% for VLA-JEPA and 47.0% for π0.5. Table-specific denominators are unexplained; binary drawer baseline values of 25% are not reconciled with Appendix A.2’s generic ten-trial description.
Go to primary source ↓e20PDF p. 9, Section 5; pp. 5–9, experiments; pp. 17–18, deployment and training
Authors identify teacher quality/coverage, short-horizon image-space priors and offline-cache cost as limitations. The evaluation reports task scores without measured deployment latency or uncertainty intervals for the retained comparisons.
Go to primary source ↓e21PDF p. 6, Figure 3(III) and caption; p. 20, Figure 7 and caption; p. 21, Figure 8
Figure 3 explicitly labels teacher-target visualizations; change peaks and the flow color wheel illustrate spatial support and motion. Figures 7–8 provide simulation prior examples and representative Dobot executions, without a quantitative causal-isolation test.
Go to primary source ↓8.5 Primary sources
Bridge-WA: Predicting Where and How the World Changes for Robotic Action ↗
PDF · 10,172 extracted words
Source fingerprint
4042f26992574ba438e03b33e897b917ce568a64a6b5c917fe12908fb87545b6