# EyeRobot

> Source: https://aiwiki.ai/wiki/eyerobot
> Updated: 2026-11-01
> Fact-checked: 2026-11-01
> Categories: Computer Vision, Embodied AI, Reinforcement Learning, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "EyeRobot." aiwiki.ai, 1 Nov 2026. https://aiwiki.ai/wiki/eyerobot
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**EyeRobot** is a line of robot-learning research from the [University of California, Berkeley](https://aiwiki.ai/wiki/uc_berkeley) that gives a manipulator robot an actively controlled "eye" and trains it to look where looking helps the arm act. The first paper, "Eye, Robot: Learning to Look to Act with a BC-RL Perception-Action Loop" (Justin Kerr, Kush Hari and colleagues, CoRL 2025), mounted a camera on a fast two-axis gimbal and trained its gaze with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning). The reward came from how well a [behavioral cloning](https://aiwiki.ai/wiki/behavioral_cloning) arm policy could predict demonstrated actions from what the eye saw.[1][2] The follow-up, "EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras" (arXiv, 2 October 2026), comes from UC Berkeley, Amazon FAR and Stanford University. It replaces the wrist cameras that most bimanual robots use with learned gaze inside a single fixed stereo camera. The authors report that it matches a standard ego-plus-wrist camera setup when the wrist views are clear (69% vs. 64% task success) and more than doubles it when grasped objects block the wrist cameras (48% vs. 22%).[5]

| | EyeRobot (1.0) | EyeRobot 2.0 |
| --- | --- | --- |
| Paper | "Eye, Robot: Learning to Look to Act with a BC-RL Perception-Action Loop"[1] | "EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras"[5] |
| arXiv | 2506.10968 (v1 12 June 2025, v2 15 September 2025)[1] | 2610.03710 (v1 2 October 2026)[5] |
| Venue | 9th Conference on Robot Learning (CoRL 2025), PMLR 305:3647-3664[2] | arXiv preprint; no venue listed on the arXiv record as of 1 November 2026[5] |
| Institutions | UC Berkeley[1] | UC Berkeley, Amazon FAR, Stanford University[5] |
| Eye | One physical camera on a two-motor pan-tilt gimbal[1] | Two virtual eye viewpoints rendered from a fixed stereo camera[5] |
| Robot | Single UR5e arm with a Robotiq gripper[1] | Bimanual 14-DoF I2RT YAM with parallel-jaw grippers[5] |
| Training environment for gaze | EyeGym: views rendered from 360-degree video recorded during teleoperation[1] | Views synthesized from calibrated stereo video[5] |
| Gaze learning | RL eye policy co-trained with a BC arm policy (BC-RL loop)[1] | Goal-conditioned gaze policy trained with RL, plus an RL target selector co-trained with a BC gripper policy[5] |
| Main comparison | Wrist, external and wrist-plus-external cameras on large-workspace tasks[1] | Passive stereo and ego-plus-wrist cameras on precise bimanual tasks[5] |
| Project page | eyerobot.net[3] | eyerobot2.github.io[6] |

## Background

### Active vision

Most robot-learning systems treat cameras as passive: the camera is fixed to the table, the robot's head or its wrist, and the policy works with whatever lands in the frame. Active perception, as set out in R. Bajcsy's 1988 paper in *Proceedings of the IEEE*, and active vision, as formalized by J. Aloimonos, I. Weiss and A. Bandyopadhyay in the first volume of the *International Journal of Computer Vision* the same year, treat sensing as something an agent does: it moves its sensors to gather the information it needs.[12][13] Both EyeRobot papers cite these works in their related-work sections.[1][5]

The EyeRobot papers are also motivated by human vision. People do not process the whole visual field at high resolution; the fovea covers a small central region, and the eyes move to put that region where it is needed. EyeRobot 2.0 cites M. F. Land and M. Hayhoe's 2001 study of eye movements in everyday activities for the observation that during tasks such as reaching for objects, human gaze moves ahead of the hand.[5][14] The first paper opens with a quotation attributed to J. J. Gibson: "We perceive in order to act and we act in order to perceive."[1]

### Active vision in robot learning

In prior and concurrent work, active vision in [imitation learning](https://aiwiki.ai/wiki/imitation_learning) mostly meant moving a camera to a better viewpoint under human control. AV-ALOHA (Chuang et al., 2024) extended the [ALOHA 2](https://aiwiki.ai/wiki/aloha_2) bimanual platform with a separate 7-DoF arm carrying a stereo camera. The operator steered it with head and body movements through a VR headset, and the robot learned the camera policy from those demonstrations.[15] Open-TeleVision (Cheng et al., 2024) let operators perceive a robot's surroundings stereoscopically while their arm and hand motions were mirrored onto humanoid robots.[16] GIAVA (Chuang et al., July 2025, a month after the first EyeRobot preprint) added human eye-tracking data and a foveated patch tokenization for [vision transformers](https://aiwiki.ai/wiki/vision_transformer), reporting lower compute and better robustness to background distractors.[17]

EyeRobot differs in where the gaze behavior comes from. It uses no gaze demonstrations; the eye is trained with RL, and its reward is tied to how well the arm policy can act.[1][5]

## EyeRobot (2025)

### Hardware

The first EyeRobot uses a "mechanical eyeball": a 90 fps, 1900x1200 global-shutter RGB camera with a 110-degree field-of-view fisheye lens, mounted on a gimbal of two direct-drive brushless DC motors that independently control pan and tilt.[1] According to the paper, the gimbal can complete a stopped-to-stopped saccade of 60 degrees in 20 ms; during inference the team limited and smoothed eye speed to reduce motion blur.[1] The eye is mounted rigidly to the base of a [Universal Robots](https://aiwiki.ai/wiki/universal_robots) UR5e arm with a Robotiq gripper.[1] The project's code repository lists the motors as LKMotor MS5005 direct-drive servo modules on an RS485 bus and includes CAD files for 3D-printing the eyeball.[4]

### EyeGym and the BC-RL loop

Training a gaze policy with RL on real hardware would require huge amounts of robot time, so the authors built EyeGym, an RL environment that simulates eye gaze by sampling views from real 360-degree video.[1] During data collection they swapped the eyeball for an Insta360 X4 360-degree camera and recorded 5.7K, 30 fps equirectangular video synchronized with robot trajectories teleoperated through the GELLO system. EyeGym can then render what the eye would have seen looking in any direction while the recorded arm motion plays back.[1]

On top of EyeGym, the paper introduces a BC-RL loop:[1]

- The **hand** is a behavior cloning agent trained to reproduce demonstrated action chunks from the rendered eye views.
- The **eye** is an RL agent (trained with [PPO](https://aiwiki.ai/wiki/ppo)) whose reward is the negative Fréchet distance between the end-effector paths of the hand's predicted and ground-truth action chunks.
- The two agents co-train: observations flow from the eye to the hand, and a task-performance signal flows from the hand back to the eye. Each episode begins with 30 paused frames so the eye can search before the demonstration advances.

Because a randomly initialized eye often cannot see the target, which makes the BC objective ill-conditioned, the authors first run Active Visual Pretraining (AVP): the eye is rewarded for centering a target object, conditioned on CLIP embeddings of the search target.[1]

### Foveal Robot Transformer

Both policies use a Foveal Robot Transformer (FoRT). Each image becomes a pyramid of centered crops at several scales, all resized to 224 pixels and embedded with a frozen [DINOv2](https://aiwiki.ai/wiki/dinov2) ViT-S encoder. The result is high resolution at the center of gaze and low-resolution context in the periphery, at a small token budget.[1] Eye actions are a categorical choice among eight azimuth-elevation directions and a zero action. The models use 3D [rotary position embeddings](https://aiwiki.ai/wiki/rotary_position_embedding) over time and image coordinates.[1]

### Experiments and results

The paper evaluates five tasks that require manipulation in an arc around the arm: eraser-on-shelf, e-stop reaching, brush handoff, screwdriver servoing and towel-in-bucket. The paper gives the workspace as 180 degrees in its introduction and 210 degrees in its experiments section.[1] The text gives 100-500 teleoperated demonstrations per task, although its dataset table lists 599 for the towel task.[1]

Camera comparisons used the same data and architecture:[1]

| Task (metric) | EyeRobot | External camera | Wrist camera | Wrist + external |
| --- | --- | --- | --- | --- |
| Eraser, pick and place (success) | 60% | 0% | 100% | 60% |
| Eraser, shelf perturbed after grasp (success) | 100% | not reported | 10% | 40% |
| E-stop, slow (success) | 100% | 100% | 80% | 100% |
| E-stop, fast (success) | 100% | 100% | 60% | 100% |
| E-stop, slow (distance to target) | 4.0 cm | 7.8 cm | 5.3 cm | 7.1 cm |
| E-stop, fast (distance to target) | 4.7 cm | 9.9 cm | 4.7 cm | 5.5 cm |

The results are mixed. A wrist camera did best when it could see the eraser, but its success fell to 10% when the shelf was moved after the grasp and the wrist camera could not search for it. The external camera often touched the eraser but never grasped it, which the authors attribute to its low resolution.[1] On other tasks the paper reports a mean screwdriver-tip error of 4.0 cm (flat target) and 5.2 cm (target tilted 45 degrees). On the brush task, the robot grasped the brush at the correct orientation in 15 of 20 trials and completed the human handover in 14 of 15.[1] In an RL-only search test with the physical eye, the policy found a towel placed outside its initial view in 87% of 15 trials, with an average search time of 1.8 seconds.[1]

The authors describe three gaze behaviors that emerged without being trained for directly: switching gaze to the next relevant object as the task progresses (for example from towel to bucket), searching for objects not yet in view, and tracking a target independently of the arm.[1] In ablations, removing foveation (a uniform 448x448 input with the same token count) raised average e-stop error from 4.4 cm to 6.4 cm and slowed settling. It also made the eye easier to distract with a yellow cup and quicker to lose a tracked object as it moved away.[1] Removing AVP lowered average towel-task success from 72.2% to 62.1%.[1]

The paper lists as limitations that 360-degree video cannot simulate motion parallax (as moving a neck would), that the policy sometimes learned "blind grasping" shortcuts reflecting gaps in the demonstration data, and that BC-RL converges much more slowly than plain behavior cloning. It suggests a second eye for stereo depth as a future direction.[1]

## EyeRobot 2.0 (2026)

### Motivation: replacing wrist cameras

EyeRobot 2.0 targets a specific design choice: the wrist-mounted cameras found on nearly all bimanual manipulation systems, such as the [ACT](https://aiwiki.ai/wiki/action_chunking_transformer)/ALOHA line.[5][18] The paper grants that wrist cameras deliver high-resolution views where manipulation happens. It argues that tying vision to the gripper is limiting when grasped objects occlude the camera and in tool use or whole-body manipulation, and that wrist cameras complicate gripper design and add motion blur.[5] The project page adds that wrist cameras make it harder to transfer from egocentric human data, which the authors call "the easiest form of manipulation data to scale."[6]

The project page also states the trade-off in token terms. A single HD image needs about 7,500 visual tokens to represent with 16x16 patches, and matching the resolution of EyeRobot 2.0's fovea with uniform resolution would take more than 10 times the tokens.[6] The authors frame their approach as information filtering rather than viewpoint seeking. The question is whether, in a workspace that is already well observed, moving the gaze can focus compute and representation capacity on the parts of the scene that matter.[6]

### Active Visual Fixation

The paper calls its method Active Visual Fixation (AVF). Despite the eye vocabulary, EyeRobot 2.0 has no moving camera: the paper describes "one fixed stereo camera" mounted between the arms.[5] Gaze is done in software. Because rotating a camera about its optical center only requires warping its image, the system renders the views two eyes would see if they rotated to converge on a shared 3D fixation point, within the wide field of view of the original cameras. The same stereo camera is used for data collection, training and deployment.[5][6]

The fixated images are processed foveally, with a pyramid of centered crops encoded by a frozen [DINOv3](https://aiwiki.ai/wiki/dinov3) ViT-S/16 backbone, so more visual tokens go to the center of gaze.[5]

### Hierarchical gaze, learned with RL

AVF splits gaze into two learned modules, both trained with RL on real-world data and neither using human gaze demonstrations:[5]

1. **Goal-conditioned gaze policy.** Given a text prompt naming a task object (in practice, from a fixed per-task list), the current fixated views and the current fixation point, it outputs changes in azimuth, elevation and fixation distance as a discretized distribution of 119 bins. It is trained with PPO on still frames from the demonstrations. The reward is the negative distance between the fixation point and the object's 3D centroid, which is computed by segmenting the object with SAM 3 and estimating depth with FoundationStereo.[5] The project page describes this step as distilling slow foundation-model object detection into a fast, real-time servoing policy, which the paper says runs at 30 Hz.[5][6] Average 3D fixation error after this pretraining was 3.5 cm.[5]
2. **Target selector.** A small MLP that reads robot proprioception and chooses which object to look at, emitting a new target every 15 frames. It is trained in a BC-RL loop as in the first EyeRobot: a BC gripper policy co-trains on the fixated views, and the selector's reward is the negative arc-length Fréchet distance between predicted and ground-truth action chunks for the "active arm" (the arm that moves more).[5]

### Fixation-centric actions

The gripper policy, which the project page describes as an ACT-style transformer decoder,[6] expresses end-effector proprioception and action chunks in a coordinate frame that rotates with the gaze. The frame is defined by the rotation of a virtual center eye toward the fixation point. It also receives the distance to the fixation point as a depth cue.[5] The authors argue that this canonicalization makes the distribution of actions more compact: reaching for the same object at different table positions looks more alike in the fixation frame than in world coordinates.[5][6]

### Hardware and training

| Item | Detail |
| --- | --- |
| Robot | Bimanual 14-DoF I2RT YAM with parallel-jaw grippers[5] |
| Camera | Static stereo camera between the arms, 35 cm above the base joint, angled 45 degrees down; each eye 120-degree diagonal FOV at 1600x1200[5] |
| Teleoperation | Custom leader arm modified from GELLO and the TRLC dev kit; a Meta Quest 3 VR interface for the simulated tasks[5] |
| Simulation | [MuJoCo](https://aiwiki.ai/wiki/mujoco) digital twin of the YAM arms, used for controlled comparisons rather than sim-to-real transfer[5] |
| Data | 1,278 demonstrations, 287.7 minutes in total, across 7 real and 6 simulated tasks (4 to 53 minutes per task in the paper's dataset table)[5] |
| Optimizers | AdamW for gaze pretraining; [Muon](https://aiwiki.ai/wiki/muon_optimizer) for the target selector and BC policy[5] |
| Compute | Each training run on 8 NVIDIA L40S GPUs[5] |
| Inference | About 20 Hz end-to-end for AVF policies, about 40 Hz for baselines[5] |

### Experiments

The real-world tasks are tape handover, pot lid placement, a six-stage toaster-oven task (open the door, pull out the rack, lift out the pan, set it down, push the rack back in, close the door), a six-stage marker task (cap a marker, insert it into a pencil pouch, zip the pouch), boba straw insertion, teabag-in-cup and opening a toolbox to extract a wrench.[5] All data was collected by the authors, and every policy was trained from scratch on the same demonstrations. Each real policy was evaluated on 25 shared test configurations per task (64 trials for tape handover), aligned by overlaying the live camera stream. In total the team reports more than 1,000 physical trials and 1,800 simulated trials.[5]

The two baselines use the same architecture and were trained with ACT. "No Gaze" sees the raw stereo stream. "Ego + Wrist" uses the left stereo image plus two wrist cameras.[5] The project page notes that both baselines see the whole table at once, which EyeRobot 2.0 does not.[6]

**Real-world end-to-end success (25 trials per task and policy, except 64 for tape handover):**[5][8]

| Task | Passive stereo (No Gaze) | Ego + Wrist | EyeRobot 2.0 |
| --- | --- | --- | --- |
| Marker | 4% | 56% | 68% |
| Tea | 28% | 80% | 80% |
| Toaster | 36% | 52% | 52% |
| Wrench | 32% | 68% | 76% |
| Boba (wrist view occluded) | 8% | 12% | 44% |
| Pot lid (wrist view occluded) | 28% | 32% | 52% |
| Mean of all seven real tasks | 27.0% | 52.2% | 66.5% |

The seven-task mean also includes tape handover, which was evaluated over 64 trials; EyeRobot 2.0 succeeded in 60 of them.[5][8] The paper's headline comparisons come from these figures. On the four tasks with clear wrist views (marker, tea, toaster, wrench), EyeRobot 2.0 averages 69% against 64% for ego-plus-wrist. On the two occluded tasks (boba and pot), it averages 48% against 22%.[5] Removing wrist cameras from a standard policy dropped mean real-world success from 52% to 27%, and EyeRobot 2.0, using the same stereo input as that passive baseline, outperformed it by what the paper reports as 40% in real and 20% in simulation, roughly the gaps between the average success rates.[5]

According to the authors, passive stereo policies tended to overfit to training positions and lacked the visual acuity for precise steps, with failures clustering at stages such as capping the marker or inserting the straw.[5] Ego-plus-wrist policies recovered grasping precision but failed at stages where the grasped object blocked the wrist cameras, such as seating the pot lid or inserting the boba straw.[5]

**Simulated tasks (100 trials each; best of 3 seeds reported):**[5][8]

| Task | Passive stereo | Ego + Wrist | EyeRobot 2.0 |
| --- | --- | --- | --- |
| Tape handover | 55% | 81% | 79% |
| Pot lid | 69% | 90% | 91% |
| Tiger pickup | 87% | 96% | 98% |
| Medical tray | 44% | 85% | 78% |
| Plate in rack | 41% | 74% | 70% |
| Hang spoon | 26% | 27% | 30% |
| Average | 54% | 75% | 74% |

In simulation, EyeRobot 2.0 roughly tied the wrist-camera baseline on average and trailed it on three of six tasks. The paper notes that simulation may favor wrist cameras compared with physical trials, and that absolute simulated success did not predict real-world success, although policy rankings held on the two tasks run in both settings.[5]

**Distractors.** With unseen colorful objects scattered on the table, the authors report the following counts out of 25 trials:[5][6]

| Policy | Wrench: first grasp / success | Tea: first grasp / success | Tape: first grasp / success |
| --- | --- | --- | --- |
| Ego only | 3 / 0 | 3 / 0 | 6 / 0 |
| Ego + Wrist | 10 / 4 | 11 / 4 | 14 / 0 |
| EyeRobot 2.0 | 20 / 5 | 13 / 4 | 23 / 8 |

The authors observed that the wrist-camera policy often reached for the largest or most distinctive object in view and could not recover once its camera lost sight of the target. EyeRobot 2.0 mostly reached for the right object when its gaze was on it, and mostly missed when it was looking in the wrong place.[5]

**Ablations.** In real-world trials, removing foveation lowered mean success from 66.5% to 46.5%, and using a single image without fixation depth (the "mono" ablation) lowered it to 48.5%.[5] In simulation, expressing actions in the world frame instead of the fixation frame dropped average success from 0.74 to 0.53. Fixing gaze on one constant object during training gave 0.50 for the worst choice and 0.69 for the best. The stereo and foveation ablations had little effect in simulation, which the authors attribute to the simulated tasks lacking visual complexity and tight tolerances.[5]

**Learned gaze.** In the marker task the target selector first fixates the marker cap, switches to the marker just before picking it up and holds it through capping. It sometimes uses static objects, such as a pencil pouch that was always roughly straight ahead during data collection, as anchors.[5] In simulation, every trained policy looked at the first object to be grasped during the first task stage.[5]

### Limitations

The paper says EyeRobot 2.0 trains only task-specific gaze and gripper policies. Moving to multi-task use would need gaze trained on a large enough corpus to accept generic prompts, possibly with a vision-language model generating targets.[5] It does not control head or neck movement, so it assumes a good head position is already available. Its object-level gaze pretraining cannot yet target parts of objects.[5]

## People and institutions

The first paper's authors are Justin Kerr, Kush Hari, Ethan Weber, Chung Min Kim, Brent Yi, Tyler Bonnen, Ken Goldberg and Angjoo Kanazawa, all listed at UC Berkeley.[1] The work was done at the AUTOLAB at UC Berkeley in affiliation with the Berkeley AI Research (BAIR) Lab. The acknowledgments credit NSF Graduate Research Fellowships for Kerr, Hari and Kim and equipment grants from NVIDIA.[1]

EyeRobot 2.0 lists ten authors.[5] Kush Hari and Justin Kerr are equal first authors ("random order"), and Ken Goldberg and Angjoo Kanazawa are equal advisors. The affiliations given in the paper are:[5]

| Author | Affiliation(s) listed |
| --- | --- |
| Kush Hari | UC Berkeley |
| Justin Kerr | UC Berkeley, Amazon FAR |
| Nidhya Shivakumar | UC Berkeley |
| Samarth Mahapatra | UC Berkeley |
| Carmelo Sferrazza | Amazon FAR |
| Jiahui Lei | UC Berkeley |
| Jitendra Malik | UC Berkeley, Amazon FAR |
| C. Karen Liu | Amazon FAR, Stanford University |
| Ken Goldberg | UC Berkeley |
| Angjoo Kanazawa | UC Berkeley, Amazon FAR |

Amazon FAR (Frontier AI & Robotics) is a team at [Amazon](https://aiwiki.ai/wiki/amazon). Amazon Science describes [Pieter Abbeel](https://aiwiki.ai/wiki/pieter_abbeel) as co-leading it.[10] Abbeel and fellow [Covariant](https://aiwiki.ai/wiki/covariant) co-founders Peter Chen and Rocky Duan joined Amazon in 2024, when Amazon also took a non-exclusive license to Covariant's robotics foundation models.[11] Abbeel is not an author of either EyeRobot paper.[1][5]

## Code, data and releases

| Release | Status (as of 1 November 2026) |
| --- | --- |
| EyeRobot code (github.com/kerrj/eyerobot) | Public. Supports BC-RL training, robot inference (implemented for a UR5 setup) and simulated eyeball-only inference on 360-degree video; includes eyeball CAD files and a link to the demonstration data; 52 stars[4] |
| EyeRobot 2.0 code (github.com/kerrj/eyerobot2) | Repository created 24 September 2026; its README says "Code and data coming soon!"; 27 stars[7] |
| EyeRobot 2.0 simulation | The paper says all simulation code will be made public, and the project page says the full simulation setup (task environments, teleoperation, training and evaluation code) will be open-sourced[5][6] |
| EyeRobot 2.0 evaluation videos | The project page hosts a browser of evaluation trial videos[6] |

## Reception

Justin Kerr announced EyeRobot 2.0 on X on 2 October 2026, opening with the question of whether wrist cameras, which "add bulk, cost, and complexity," are really needed. As of 1 November 2026 the post showed about 106,000 views and 459 likes.[9]

## See also

- [Embodied AI](https://aiwiki.ai/wiki/embodied_ai)
- [Robot manipulation](https://aiwiki.ai/wiki/robot_manipulation)
- [Robot learning](https://aiwiki.ai/wiki/robot_learning)
- [Robot teleoperation](https://aiwiki.ai/wiki/robot_teleoperation)
- [Universal Manipulation Interface](https://aiwiki.ai/wiki/universal_manipulation_interface)

## References

1. Kerr, J.; Hari, K.; Weber, E.; Kim, C. M.; Yi, B.; Bonnen, T.; Goldberg, K.; Kanazawa, A. ["Eye, Robot: Learning to Look to Act with a BC-RL Perception-Action Loop"](https://arxiv.org/abs/2506.10968). arXiv:2506.10968 (v2, September 15, 2025).
2. Proceedings of Machine Learning Research. ["Eye, Robot: Learning to Look to Act with a BC-RL Perception-Action Loop"](https://proceedings.mlr.press/v305/kerr25a.html). Proceedings of The 9th Conference on Robot Learning, PMLR 305:3647-3664, 2025.
3. EyeRobot project page. ["Eye, Robot: Learning to Look to Act with a BC-RL Perception-Action Loop"](https://www.eyerobot.net/). Accessed November 1, 2026.
4. Kerr, J. ["kerrj/eyerobot: Source code for EyeRobot"](https://github.com/kerrj/eyerobot). GitHub repository. Accessed November 1, 2026.
5. Hari, K.; Kerr, J.; Shivakumar, N.; Mahapatra, S.; Sferrazza, C.; Lei, J.; Malik, J.; Liu, C. K.; Goldberg, K.; Kanazawa, A. ["EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras"](https://arxiv.org/abs/2610.03710). arXiv:2610.03710 (v1, October 2, 2026).
6. EyeRobot 2.0 project page. ["EyeRobot 2.0: Active Gaze for Precise Manipulation without Wrist Cameras"](https://eyerobot2.github.io/). Accessed November 1, 2026.
7. Kerr, J. ["kerrj/eyerobot2: Code release for EyeRobot 2.0"](https://github.com/kerrj/eyerobot2). GitHub repository. Accessed November 1, 2026.
8. EyeRobot 2.0 project page. ["Results chart data"](https://eyerobot2.github.io/data/results-charts.json). Per-task success values behind the project page's charts. Accessed November 1, 2026.
9. Kerr, Justin (@justkerrding). ["Almost all bimanual robots include wrist cameras..."](https://x.com/justkerrding/status/2106085571311579357). X, October 2, 2026.
10. Amazon Science. ["Pieter Abbeel"](https://www.amazon.science/author/pieter-abbeel). Author profile. Accessed November 1, 2026.
11. SiliconANGLE. ["Amazon hires founders of AI robotics startup Covariant"](https://siliconangle.com/2024/09/02/amazon-hires-founders-ai-robotics-startup-covariant/). September 2, 2024.
12. Bajcsy, R. ["Active perception"](https://doi.org/10.1109/5.5968). *Proceedings of the IEEE* 76(8):966-1005, 1988.
13. Aloimonos, J.; Weiss, I.; Bandyopadhyay, A. ["Active vision"](https://doi.org/10.1007/BF00133571). *International Journal of Computer Vision* 1(4):333-356, 1988.
14. Land, M. F.; Hayhoe, M. ["In what ways do eye movements contribute to everyday activities?"](https://doi.org/10.1016/S0042-6989(01)00102-X). *Vision Research* 41(25-26):3559-3565, 2001.
15. Chuang, I.; Lee, A.; Gao, D.; Naddaf-Sh, M.; Soltani, I. ["Active Vision Might Be All You Need: Exploring Active Vision in Bimanual Robotic Manipulation"](https://arxiv.org/abs/2409.17435). arXiv:2409.17435, 2024.
16. Cheng, X.; Li, J.; Yang, S.; Yang, G.; Wang, X. ["Open-TeleVision: Teleoperation with Immersive Active Visual Feedback"](https://arxiv.org/abs/2407.01512). arXiv:2407.01512, 2024.
17. Chuang, I.; Zou, J.; Lee, A.; Gao, D.; Soltani, I. ["Look, Focus, Act: Efficient and Robust Robot Learning via Human Gaze and Foveated Vision Transformers"](https://arxiv.org/abs/2507.15833). arXiv:2507.15833, 2025.
18. Zhao, T. Z.; Kumar, V.; Levine, S.; Finn, C. ["Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware"](https://arxiv.org/abs/2304.13705). arXiv:2304.13705, 2023.