ActiveGlasses: Learning Manipulation with Active Vision from Ego-centric Human Demonstration
1. Paper overview
In one sentence: ActiveGlasses learns object and camera motion from bare-hand demonstrations, trading direct hand retargeting for calibrated 3D perception and task-specific trajectory learning. e-framinge-processinge-methode-resultse-ablation
| At a glance | What to know |
|---|---|
| Research problem | Source description Bare-hand demonstrations preserve natural manipulation, but human and robot kinematics differ. Head motion also changes image coordinates and reveals otherwise occluded targets. The paper asks whether object-centered supervision and a separately actuated camera can transfer both behaviors without robot-specific policy finetuning. e-identitye-framing |
| Core mechanism | Source description A glasses-mounted stereo capture interface records bare-hand manipulation and head motion, with gesture-controlled episodes, audio feedback and synchronized timestamps. e-hardware |
| A key reported result | Book placement: 14/20 (70%) Final task successes / trials. Physical evaluation; 200 demonstrations; 20 trials per method; randomized shelf pose and empty slot. Stage 3 requires collision-free placement. Without active vision: 7/20; π0.5: 7/20. A derived 35-percentage-point improvement over either comparator. Intermediate ActiveGlasses counts are 20/20 and 16/20. e-taskse-protocole-results |
| Reading caution | Reader analysis The π0.5 comparator uses Sigma-collected real-robot demonstrations and modified camera inputs; it is not a pure representation ablation. Evaluation randomization ranges, repeated training seeds and statistical uncertainty for task success are not supplied. Three tabletop tasks do not establish unrestricted generalization. e-protocole-results |
Core contributions
- Source description
A glasses-mounted stereo capture interface records bare-hand manipulation and head motion, with gesture-controlled episodes, audio feedback and synchronized timestamps. e-hardware
- Source description
A RISE-derived point-cloud policy separates absolute object-trajectory prediction from relative camera-motion prediction, enabling demonstrated transfer between Flexiv and UR5. e-methode-transfer
Figure 2. Human and robot observations meet in a shared spatial representation. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the orange collection branch: stereo images and tracked head motion support depth, pose estimation and calibration. The purple deployment branch retains real-time stereo processing and calibration. Follow the central point-cloud block through the 3D transformer and action token, then follow the two blue outputs toward different physical arms. The large return arrow comes from the perception arm to the observation side, indicating that moving the camera changes subsequent input. The overview groups ground-truth object/head poses beside RGB and depth; Sections III-B and III-C distinguish demonstration supervision from the final policy's point-cloud input and explicitly omit an extra current-object-pose condition. e-overviewe-processinge-methode-rolloute-policy-figure
What it supports. The key interface is spatial: human demonstrations supply desired object movement without requiring a mapping from every human joint to robot joints. Camera movement is also learned and physically executed. The feedback arrow supports an observation–action loop, while Algorithm 1 supplies the actual sequence of reconstruction, prediction and execution.
Where the evidence stops. This overview draws a single Diffusion Head block. Section III-C and Figure 6 instead specify two separate diffusion heads, which this report follows. The grouped ground-truth boxes must not be read as proof that the deployed default policy receives ground-truth poses.
2. Motivation
2.1 The problem and the proposed response
Bare-hand demonstrations preserve natural manipulation, but human and robot kinematics differ. Head motion also changes image coordinates and reveals otherwise occluded targets. The paper asks whether object-centered supervision and a separately actuated camera can transfer both behaviors without robot-specific policy finetuning. e-identitye-framing
2.2 What this reading follows
A useful demonstration contains more than a hand trajectory: it also shows where the demonstrator looks. ActiveGlasses records both with a stereo camera mounted on tracked glasses, then deploys the camera on a separate robot arm. The learning interface is a world-frame point cloud, and the manipulation target is the object's trajectory. This edition follows the conversion from human recording to robot execution, then separates task success, policy ablations and data-collection efficiency. The experiments support active perception in three constrained tabletop tasks. They leave open how reliably the calibration, training recipe and visual responsiveness transfer beyond the reported settings. e-framinge-processinge-methode-resultse-ablation
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 Datasets. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The foundational data-collection/action-policy classification is supported. This architecture learns desired object and sensing actions from observations; it does not specify an action-conditioned environment transition model or imagined-future planning. Two action diffusion heads do not establish either a One Model world-action architecture or joint future-world/action prediction. The recorded world-model quadrant remains not applicable. e-methode-rolloute-policy-figure
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 First make camera movement compatible with stable geometry
The same shelf can occupy very different image coordinates when the demonstrator leans or turns. ActiveGlasses reconstructs stereo depth and places the resulting point cloud into a common world frame so those viewpoint changes need not redefine the task coordinates. Three spheres initialize that frame, and head tracking propagates it after initialization. During demonstration processing, hand masking removes human-specific points and object pose estimation creates supervision. A mesh prior helps that estimation; bare-hand collection therefore does not mean annotation-free learning. Reader analysis: this interface makes calibration quality a prerequisite for the claimed spatial consistency. The paper's inverse-transform wording and index changes between Equations (2)–(4) and Algorithm 1 remain unresolved, so those formulas need geometric verification before becoming implementation instructions. e-processinge-calibratione-methode-rollout
Figure 6. Separate output heads let object and camera motion use different coordinates. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the shared backbone from left to right, then separate the upper manipulation branch from the lower perception branch. The dashed legend labels ablation variants, so the dashed current-object-pose box is an experimental input choice rather than a mandatory part of the default system. The manipulation output box similarly tests absolute versus relative object trajectories. The lower branch retains relative head motion. Cross-checking Section III-C establishes the default: absolute object targets, no explicit current-pose conditioning, and relative camera movement. This figure clarifies the two-head architecture that the broader system overview compresses into one block. e-policy-figuree-methode-heade-ablation
What it supports. Objects and cameras have different transfer constraints. Absolute object targets match the policy's world-frame geometry, while relative head targets reduce dependence on the demonstrator's initial head position. The authors report that absolute head targets can demand large initial motions and bring the perception arm near its workspace limits.
Where the evidence stops. The drawing explains architectural choices, not their independent effect sizes. The paper raises single-versus-separate-head design as an experimental question but supplies no numerical comparison. Its absolute-head failure discussion is qualitative and should not be converted into a success-rate claim.
5.2 Then separate the desired object motion from the robot that executes it
The policy's manipulation output describes where the object should move, rather than directly imitating the demonstrator's finger joints. A pre-grasp procedure establishes the grasp, and calibration connects object targets to end-effector motion. A second diffusion head predicts relative head movement so the perception arm can obtain useful views without copying the human's absolute starting position. These two outputs share point-cloud features but have different coordinate conventions. Algorithm 1 returns to stereo reconstruction during rollout and also predicts a termination flag trained from episode endings. Reader analysis: this is an observation-driven action policy with active sensing. Predicting a desired object trajectory does not by itself learn how the environment will evolve under candidate actions; no separate future-world simulation or rollout-based action search is specified. e-methode-rolloute-heade-policy-figure
5.3 Finally distinguish a winning configuration from an isolated mechanism
Table I first establishes an operational result: the complete system finishes more trials than its fixed-camera variant on every task. That comparison holds the backbone and action representation constant, making it particularly relevant to active vision. The π0.5 comparison uses a different demonstration pipeline and representation, so its gains have a broader interpretation. Table II then tests design choices inside the proposed system. Adding current pose harms absolute-output success, but relative output with current pose outperforms absolute output with current pose. The absent relative/no-pose result prevents a complete factorial conclusion. Reader analysis: the default is empirically supported, whereas the stronger explanation that extra pose information causes visual neglect remains a hypothesis requiring a targeted observation intervention. Table III separately tests transfer, with a remaining final-stage performance gap. e-protocole-resultse-ablatione-transfer
5.4 Training and inference
During training
The policy learns from processed human demonstrations; Section IV-B describes training from scratch. Table I gives 200 book-placement, 100 bread-insertion and 100 pouring demonstrations. The last five frames of each episode receive termination label 1; earlier frames receive 0. No explicit diffusion loss, optimizer, training duration or frozen-module schedule is supplied. e-methode-protocole-results
During inference
Algorithm 1 repeatedly estimates depth, reconstructs and aligns the cloud, clips distant regions, predicts both trajectories and executes them. It stops when the predicted termination flag exceeds a threshold. The numerical threshold, action horizon, execution frequency and precise replanning schedule are unspecified. Relative head motion is chosen to reduce workspace-limit and inverse-kinematics failures caused by human–robot starting-pose differences. e-rolloute-head
5.5 Implementation flow
- Capture and reconstruct
XREAL Air 2 Ultra tracks the head; a mounted ZED Mini supplies stereo images because direct access to XREAL camera streams is unavailable. FoundationStereo estimates depth. During demonstration processing, Grounded-SAM removes human-hand points, SAM2 masks the object, and FoundationPose estimates its poses using an object mesh as a geometric prior. e-hardwaree-processing
- Stabilize coordinates
Hand–eye calibration relates glasses and camera. Three orange tabletop spheres establish the initial world frame; subsequent head-relative motion propagates alignment when the spheres become occluded. This gives the policy spatially consistent observations despite camera movement. The printed transform conventions need clarification before implementation. e-calibratione-method
- Predict and execute
AnyGrasp handles pre-grasp, with a fixed strategy for tasks requiring precise grasp poses. The learned stage predicts object and head trajectories with two diffusion heads. The default manipulation head omits explicit current-object-pose conditioning. A calibrated object-to-end-effector transform converts object targets into robot actions; trajectory prediction alone is not execution. e-methode-rolloute-policy-figure
6. Experiments & results
ActiveGlasses transfers bare-hand demonstrations to robots by learning object motion and camera motion from a common 3D representation. Human stereo video supplies geometry and object-trajectory supervision; a robot arm then moves the same camera while another manipulates the object. Its contribution is an imitation-learning and data-collection system, with measured physical-task success rather than generated-video evaluation. The strongest controlled evidence favors active vision within the authors' policy; comparisons with π0.5 also change representation and demonstration collection.
6.1 Read the original evidence
Table I. Active vision improves final completion, with substantial failures remaining. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each task as a three-stage progression, using Stage 3 for final task success. These are counts out of the original twenty trials, not conditional probabilities between stages. The tasks use 200, 100 and 100 demonstrations respectively; evaluation randomizes the shelf and slot, toaster, or cup and screen within unspecified predefined ranges. The no-active-vision row keeps the policy backbone and action representation. The Pi05 row refers to π0.5 trained on Sigma-collected robot demonstrations with head imagery as its only visual input. The paper's task descriptions define collision-free book placement, completed bread insertion and water poured into the cup as the final outcomes. e-taskse-protocole-results
What it supports. ActiveGlasses achieves final success rates of 70%, 55% and 50%. Subtracting π0.5's rates gives gains of 35, 25 and 30 percentage points. Intermediate progress can mislead: π0.5 reaches bread Stage 2 in 18/20 trials versus 15/20 for ActiveGlasses, yet finishes fewer insertions.
Where the evidence stops. The π0.5 comparison changes representation and demonstration collection, so it does not isolate either factor. Twenty trials per method and no reported repeated-seed uncertainty limit precision. The reported gains concern physical task completion in this setup, not general manipulation competence.
Figure 5. Collection efficiency and hand burden are separate from learned-policy success. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The upper panel has two scales: bar heights use completion time in seconds on the left, while the orange line uses success percentage on the right. Within each task the method order follows the colored bars. These measurements concern collecting demonstrations, whereas Table I concerns the trained robot policy. The lower panel compares reported device burden and adds context labels. ActiveGlass is labeled Bare Hand; Sigma is labeled Limited Workspace, and the caption explains its zero-gram entry through zero-gravity operation. Read these annotations with the bars rather than interpreting them as a weighing of every component in each complete system. e-collectione-hardwaree-results
What it supports. The plotted ActiveGlass collection times are lower than those of the other interfaces on both tasks, while its collection-success line reaches the top of the percentage scale. This supports the authors' efficiency motivation. The zero-gram bare-hand entry describes the absence of a handheld device; Figure 3 separately shows nonzero camera and glasses masses.
Where the evidence stops. The source does not provide participant counts, a full collection-study protocol or definitions for the displayed error bars. The zero-gram Sigma and ActiveGlass entries are contextual conventions, not total-system mass. No precise timing values are inferred from unlabeled bar heights here.
Table III. A shared object target transfers, while execution constraints remain arm-specific. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the two robot rows at the same stage. Both arms approach the shelf successfully in every trial and reach insertion in sixteen trials, but their collision-free completion counts diverge. Section IV-D says the policy is deployed on UR5 in the same setting. Section III-C explains why this is possible: the learned target describes the object's pose, and calibration converts it into end-effector motion. This reduces dependence on the training demonstrator's embodiment, but it does not remove the new arm's reachable-workspace constraints or the need to establish a suitable grasp and calibrated object-to-gripper relationship. e-transfere-methode-tasks
What it supports. UR5 completes 11/20 placements compared with 14/20 for Flexiv, despite matching the first two stage counts. The result demonstrates transfer of the task policy to another physical arm. It also locates the measured degradation at final placement, where geometry and execution constraints become especially consequential.
Where the evidence stops. The authors attribute the UR5 failures to its smaller workspace, but no workspace-matched control isolates that explanation. This table evaluates book placement only; it does not establish equal transfer performance for bread insertion, pouring or arbitrary robot embodiments.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Book placement Physical evaluation; 200 demonstrations; 20 trials per method; randomized shelf pose and empty slot. Stage 3 requires collision-free placement. | 14/20 (70%) Final task successes / trials | Without active vision: 7/20; π0.5: 7/20. A derived 35-percentage-point improvement over either comparator. Intermediate ActiveGlasses counts are 20/20 and 16/20. e-taskse-protocole-results |
| Bread insertion Physical toaster task; 100 demonstrations; 20 trials per method; randomized toaster pose; slot initially invisible. | 11/20 (55%) Final task successes / trials | Without active vision: 0/20; π0.5: 6/20. A derived 25-percentage-point gain over π0.5. Its stronger Stage-2 count, 18/20 versus 15/20, does not translate into better final insertion. e-taskse-protocole-results |
| Occluded distant water pouring Physical pouring behind a screen; 100 demonstrations; 20 trials per method; randomized cup and screen poses. | 10/20 (50%) Final task successes / trials | Without active vision: 4/20; π0.5: 4/20. A derived 30-percentage-point gain over either comparator, while half the ActiveGlasses trials still fail. e-taskse-protocole-results |
| Book-placement action representation ablation Table II; collision-free completion; 20 trials for each reported configuration. | Absolute output without current-pose conditioning: 14/20 Final task successes / trials | Absolute with current pose: 3/20; relative with current pose: 10/20; relative without current pose: dash, not a reported score. Supports the default configuration, but the incomplete factorial table cannot establish an independent universal advantage of absolute coordinates. e-ablation |
| Book-placement transfer from Flexiv to UR5 Table III; same task setting, policy deployed on a second manipulation arm; 20 trials per arm. | UR5: 11/20 Final task successes / trials | Flexiv Rizon 4: 14/20; both achieve 20/20 and 16/20 in Stages 1 and 2. Demonstrates transfer on one task. The authors attribute additional final-stage failures to UR5 workspace limits; no isolated causal test is reported. e-transfer |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table II. The default wins, but the missing cell limits the mechanism conclusion. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Rows select the object trajectory's coordinate representation; columns select whether the manipulation diffusion head receives the current object pose as an extra condition. The bold upper-left entry is the deployed default, and success requires placement without collision. Compare horizontally in the absolute row to assess the reported effect of adding current pose. Compare vertically in the with-pose column to see that relative output actually performs better under that condition. The lower-left dash has no numerical interpretation supplied by the paper. Figure 6 locates both interventions on the manipulation branch; neither table axis changes the relative-head design. e-ablatione-policy-figuree-method
What it supports. Absolute prediction without the added pose condition yields 14/20 successes, versus 3/20 when that condition is added. Relative prediction with pose reaches 10/20. This supports the particular default combination, while the with-pose comparison prevents a blanket claim that absolute coordinates always outperform relative ones.
Where the evidence stops. The authors interpret pose conditioning as encouraging near-fixed trajectories and weaker visual dependence. Table II measures success, not attention or causal visual responsiveness. The dash is an unreported result, and the source does not say whether that variant was infeasible or simply omitted.
7. Analysis & limitations
7.1 What the evidence leaves open
The π0.5 comparator uses Sigma-collected real-robot demonstrations and modified camera inputs; it is not a pure representation ablation. Evaluation randomization ranges, repeated training seeds and statistical uncertainty for task success are not supplied. Three tabletop tasks do not establish unrestricted generalization. e-protocole-results
Figure 2 compresses prediction into one diffusion-head block; Section III-C and Figure 6 specify two. Its grouped ground-truth poses should not be interpreted as required default inference conditions. Calibration notation also conflicts: the transform called camera-to-world is inverted in Equation (4), while Algorithm 1 changes index ordering and applies its transform directly. e-overviewe-calibratione-methode-rolloute-policy-figure
Figure 5 supports a collection-efficiency trend, but participant counts and error-bar definitions are absent. Its zero-gram ActiveGlasses label concerns bare hands, not headset mass. No quantitative single-versus-two-head comparison or absolute-head ablation score is reported. e-hardwaree-collectione-ablatione-head
7.2 Questions for discussion
- Does pose conditioning reduce response to a changed target when the grasped object's pose is held fixed? [e-ablation]
- How much transfer failure remains after matching reachable workspaces across arms? [e-transfer]
8. Reproducibility audit
8.1 Requirements and known gaps
Required components include synchronized stereo/head streams, object meshes, sphere and hand–eye calibration, the named perception modules, a RISE-derived policy, and an object-to-gripper transform. The evaluated hardware is Flexiv Rizon4 with Robotiq 2F-85, I2RT YAM perception arm, XREAL glasses and ZED Mini. Compute, software versions, model configuration, point-cloud clipping bounds and robot-control details remain unspecified. e-hardwaree-processinge-calibratione-methode-rolloute-tasks
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: Test whether current-pose conditioning reduces visual responsiveness
Reader-proposed, not performed: train the two absolute-output variants on identical demonstrations with matched compute and repeated seeds, changing only the extra current-object-pose input. At evaluation, hold the grasped book pose and camera pose fixed while moving the target slot within a reachable range. Compare predicted object trajectories, collision-free completion and response to the changed target; include an unchanged-target control. The shortcut explanation predicts that the pose-conditioned variant will adapt less to the visible slot change. Comparable adaptation would weaken that explanation even if its overall success remains lower. Use matched visible scenes to separate visual responsiveness from grasp or reachability failures. e-methode-taskse-ablation
Check 2: Verify the world frame before testing the learned camera policy
Reader-proposed, not performed: keep the spheres and several held-out tabletop points stationary while moving the tracked camera through translated and rotated views. Implement explicit transform directions consistent with each reading of the printed calibration notation, and compare world-coordinate drift, sphere-origin placement and camera-to-world-to-camera round-trip error. Use a fixed-camera control and fresh sphere-based calibration as references, then repeat when spheres leave view and alignment depends on head tracking. A valid convention should preserve the static geometry across views; systematic motion or offsets would falsify that implementation. Report sensitivity to nonorthogonal sphere placement rather than silently adding an unstated correction. e-hardwaree-calibratione-rollout
8.3 Reading coverage
Visual audit: All eight supplied PDF pages were rendered and visually inspected, including the title/authors/version on p. 1; framing on p. 2; Figures 2–3 and processing on p. 3; calibration equations and algorithm design on p. 4; Algorithm 1, hardware, tasks and comparator setup on p. 5; Figure 4 and Tables I–II on p. 6; Figures 5–6, Table III, head-motion discussion and conclusion on p. 7; and the final references on p. 8. All six final original crops were individually inspected, including the corrected right edge of Figure 6. The Figure 2 single-block versus two-head discrepancy and grouped supervision/input ambiguity were checked against Section III-C and Figure 6. Table II's dash and Figure 5's weight conventions are preserved. All four text chunks were read in full. No separate supplement, external code or experimental run was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
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
- I. Introduction
- II. Related Works
- III. The ActiveGlasses System: A. Hardware and Interface Design; B. Data Processing; C. Algorithm Design
- IV. Experiments: task design and hardware; A. Scalability; B. Active Vision; C. Policy Design; D. Cross Embodiment
- V. Conclusion
- References [1]–[41]
Outside the original text pass
- Identity: the title and all eight authors match the supplied catalog. The inspected title page identifies arXiv:2604.08534v1 [cs.RO], 9 April 2026. No alternative edition or revision history was supplied; no version comparison is claimed.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- That extraction limitation was addressed by visually inspecting all eight supplied PDF pages and all six final crops.
- Separate supplemental material availability has not been fully verified.
- No separate supplement or appendix was supplied; the eight-page PDF ends with references. Code and external resources were not inspected, and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/author block, arXiv margin and Abstract
Exact title, eight authors and three affiliations; arXiv:2604.08534v1, 9 April 2026; bare-hand active-vision transfer objective.
Go to primary source ↓e-framingPDF pp. 1–2, Section I
Motivates natural manipulation and independent sensing; proposes object-centered transfer with a separate perception arm.
Go to primary source ↓e-overviewPDF p. 3, Figure 2 and caption
Capture/deployment branches, 3D transformer, action token, single drawn diffusion-head block, two motion outputs and camera feedback arrow.
Go to primary source ↓e-hardwarePDF p. 3, Section III-A and Figure 3
XREAL tracking, ZED stereo, Unity interface and ROS alignment. Figure 3 labels the camera 60 g and glasses 83 g; ZED trajectory jumps motivate glasses tracking.
Go to primary source ↓e-processingPDF pp. 3–4, Section III-B.1–2
FoundationStereo depth, Grounded-SAM hand removal, SAM2 object masks and FoundationPose object trajectories with mesh prior.
Go to primary source ↓e-calibrationPDF p. 4, Section III-B.3, Equations (1)–(4)
Three-sphere axes and initial origin; hand–eye calibration and head-relative propagation. Equation (4) applies an inverse despite the camera-to-world wording.
Go to primary source ↓e-methodPDF p. 4, Section III-C
AnyGrasp/fixed pre-grasp, RISE-derived point-cloud policy, two diffusion heads, absolute object and relative head output, omitted current-pose condition, calibrated action conversion and last-five-frame termination labels.
Go to primary source ↓e-rolloutPDF p. 5, Algorithm 1
Depth, cloud reconstruction, initial/propagated calibration, direct world transformation, distant-region clipping, policy call, execution and termination threshold; numeric settings are absent.
Go to primary source ↓e-tasksPDF p. 5, Task Design and Hardware Setup; p. 6, Figure 4 and caption
Three tasks and stage criteria; manipulation and perception hardware; perception-arm base randomized within an unspecified small range.
Go to primary source ↓e-protocolPDF p. 5, Section IV-B; p. 6, Table I caption
No-active-vision variant retains backbone/action representation. π0.5 uses robot demonstrations with head imagery and modified wrist inputs. Task demonstration counts and object randomization are stated.
Go to primary source ↓e-resultsPDF p. 6, Table I, all task/stage columns and caption
Final successes: ActiveGlasses 14,11,10; no active vision 7,0,4; Pi05 7,6,4; all denominators 20. Training demonstrations: 200,100,100.
Go to primary source ↓e-ablationPDF pp. 5–7, Section IV-C; p. 6, Table II
Absolute/no-pose 14/20; absolute/pose 3/20; relative/pose 10/20; relative/no-pose dash. Authors discuss visual shortcutting and relative-output tracking failures.
Go to primary source ↓e-policy-figurePDF p. 7, Figure 6 and caption
Shared object-centric RISE features feed separate manipulation/perception diffusion heads; dashed marks denote conditioning and output-representation ablations.
Go to primary source ↓e-headPDF p. 7, Section IV-C continuation in bottom-left and top-right columns
Absolute head targets reportedly cause large initial movement and workspace/IK failures; no numerical head-representation comparison is supplied.
Go to primary source ↓e-collectionPDF p. 5, Section IV-A; p. 7, Figure 5 and caption
Collection time bars, success lines and device-weight panel. Sigma zero grams denotes zero-gravity operation; ActiveGlass zero grams is labeled Bare Hand. Participant counts and error-bar definitions are not given.
Go to primary source ↓e-transferPDF p. 7, Section IV-D and Table III
Flexiv stage counts 20/20,16/20,14/20; UR5 20/20,16/20,11/20. Authors attribute later UR5 failures to its smaller workspace.
Go to primary source ↓8.5 Primary sources
ActiveGlasses: Learning Manipulation with Active Vision from Ego-centric Human Demonstration ↗
PDF · 6,300 extracted words
Source fingerprint
cbb3f28f6681b01fc87e6c24b2f78981ac45a19b6a6fb4d5e419532930861b68