Retain the evidence.
Task-guided history inversion preserves source evidence and visible changes while transforming appearance during training.
WORLD–ACTION MODELS / ACTIVE VISION
ActiveWAM.
Evidence-Aware Active Vision
for World-Action Models
Learn what to keep.
Decide where to look.
A unified world–action model that connects task-guided history learning with physical head and arm control.
arXivResearch paperComing soon
GitHub ↗ActiveWAM codeComing soon
RoboTwin-AV ↗Benchmark repositoryView repository
REAL ROBOTPAN +36.2° · TILT +40.2°
REAL HEAD MOTION · ORIGINAL SPEEDA changing view.
↗A changing view.
A continuing task.
50RoboTwin-AV benchmark tasks
16-DBimanual + pan/tilt actions
3Physical kitchen tasks
Raw RGBAt deployment · inversion during training
01 / THE IDEA
Seeing more
is only half the story.
Moving the camera can reveal a useful relation. It can also move a critical cue out of view.
ActiveWAM treats active vision as a retain–acquire problem: learn which evidence should survive visual changes, then jointly decide how to move the head and hands. A recent, view-aware observation history connects the two.
Read the abstract ＋
Active vision manipulation requires a policy to control both its camera and its end-effectors, yet camera motion determines which evidence remains visible within finite observation windows. Acquiring a new view can displace task-critical cues, while retaining a view forgoes potentially useful observations. We formulate this as an evidence-aware retain–acquire problem and present ActiveWAM, a unified world–action model that learns observation and manipulation jointly. Training-time inversion constrains a frozen video prior by task-bearing source evidence and visible temporal changes. At deployment, the policy generates bimanual and pan/tilt actions from view-aware raw history and updates context from newly measured RGB observations. Future-video prediction serves as a co-training signal. RoboTwin-AV extends 50 manipulation tasks with executable pan/tilt control and automatically generated demonstrations.
Acquire the next view.
Head motion and bimanual manipulation share one generator. Staying still is a valid observation action.
Act on what was seen.
Execute a short action prefix, acquire actual RGB, and rebuild context from the available observation window.
02 / IN ACTION
Follow the task.
Watch the view change.
Explore real RGB, source-aligned head poses, and key moments across three domains.
ROBOTWIN-AV / SOURCE DEMONSTRATION
Bread into the skillet
HEAD ORIENTATION SOURCE ALIGNED
Head telemetry is unavailable for this recording.
PAN—
TILT—
RGB + source-aligned pan/tilt · original playback timing
These sequences visualize source demonstrations and camera motion.
Download sequence ↓03 / THE METHOD
One model.
Head and hands together.
Learning to preserve task evidence and learning to control observation, within a unified world–action model.
TRAINING ONLY
Transform appearance. Preserve task evidence.
A frozen Wan video prior transforms each camera’s observed history independently. Task-weighted source preservation and temporal-change constraints guide partial inversion. Raw and accepted transformed histories share the same original action and future-video targets.
FIG. 02 / ARCHITECTURE
Inversion is used during training. Deployment conditions on raw history and returns one joint action trajectory; future-video prediction is a co-training signal.
04 / INSIDE THE VISUAL HISTORY
Change the appearance.
Keep the task in sight.
Explore our recent visualization studies on the six key stages of a real cooking sequence.
ILLUSTRATIVE PREVIEWTask: scramble eggs
SourceStronger
These previews use designed RGB perturbations to illustrate a visual target. The levels are display settings, not diffusion timesteps or model measurements.
View the full strength sweep ↗
Archived design study. The original “tau” labels denote synthetic noise multipliers, not the paper’s inversion parameter.
View the model-produced TAVIS prototype ↗
Archived partial-inversion cache. Raw/transformed pairs are shown for head, left wrist and right wrist, with three history timestamps per row. This is a prototype, distinct from the illustrative real-robot previews above.
RoboTwin appearance studies 5 task sequences +
Design studies from the previous visualization iterations; image post-processing, not checkpoint outputs.
05 / EXPERIMENTS
Evidence across
views and environments.
Comparing manipulation success across visual and camera-pose shifts. Select a benchmark and condition to explore the results.
COMPOUND SHIFT
53.3%
Success under appearance + head-pose shifts.
50 tasks · 3 seeds · 100 episodes per task, seed and condition.
+20.0 pp vs. Fast-WAMSUCCESS RATE (%) ↑
050100
View exact values & evaluation protocol ＋
Compound success over the raw-history control (53.3% vs. 41.7%).
Head travel per episode for joint learned control, compared with 4.10 rad for object tracking.
Full-task successes across three physical kitchen tasks; 20 attempts per task.
Manuscript-reported results; evaluation protocols and exact values are available above. The gallery presents source demonstrations separately from the scored evaluations.
06 / GO DEEPER
A new perspective
on robot learning.
Citation
Anonymous manuscript · citation draft@misc{activewam2026,
title = {ActiveWAM: Evidence-Aware Active Vision
for World-Action Models},
author = {Anonymous Authors},
year = {2026},
note = {Manuscript}
}