Active Vision

Robots manipulating objects in clutter must decide not only how to act, but also where to look. Although active camera motion can expose an occluded target, it may move the robot into an initial observation-state distribution that differs from the manipulation policy's training distribution. We identify this problem as first-frame distribution sensitivity: under perturbed starts in LIBERO, naively prepending camera repositioning to pi_0.5 reduces success from 46.3% to 11.5%. We introduce ActiveAlign, a data-adaptation framework that trains a Vision-Language-Action (VLA) policy from the states at which active perception terminates. A learned viewpoint module first generates a camera trajectory. Offline, ActiveAlign uses closed-loop alignment to connect the trajectory endpoint to an existing demonstration, replays the remaining manipulation trajectory, and retains successful continuations for fine-tuning. At deployment, the same viewpoint module precedes the adapted VLA without modifying its architecture. Across 40 LIBERO tasks, ActiveAlign achieves 70.7% success with pi_0 and 72.7% with pi_0.5, outperforming their standard-training baselines by 40.0% and 26.4% respectively. Matched-state and data-matched controls support gains beyond prefix shortening and generic viewpoint augmentation; OpenVLA and GR00T N1 also show the same trend. Success rate under physical occlusion in real world increases from 35/90 to 83/90. These results show that active perception benefits manipulation when policy training is aligned with the observation and robot-state distribution induced by viewpoint selection.

Embodied Visual Tracking (EVT) is an essential capability for intelligent robots in human-robot collaboration, requiring an agent to follow a designated target person in dynamic environments relying solely on egocentric visual perception. However, persistent pedestrian occlusions in dynamic scenes pose critical challenges to existing EVT methods. Lacking proactive visibility preservation, reactive tracking methods frequently experience prolonged target loss under severe occlusions. Therefore, we propose VP-EVT, a framework that treats target visibility as an explicit objective of motion control rather than a by-product of following. Acting on this objective requires understanding both what obstructs the target and how robot motion can change that obstruction. VP-EVT therefore uses language-guided perception to extract separate target and pedestrian observations. When occlusion makes current observations incomplete, VP-EVT combines visual history with ego-motion to maintain estimates of their relative positions and motion. Using these estimates and the current following waypoints, VP-EVT predicts whether continued following will sustain the obstruction. When the predicted risk is low, the original following motion is retained; otherwise, a recurrent controller combines updated estimates with maneuver context from preceding decisions to coordinate viewpoint adjustment and the return to stable following. Moreover, we introduce OccEVT-Bench for closed-loop evaluation of persistent occlusion and collect expert demonstrations for training. On its Risk split, VP-EVT improves success rate by 121.4% relative to a competitive baseline and reduces visibility deficit and longest target-loss duration by 78.8% and 74.4%, respectively.

Inspired by human vision, we introduce a framework using active gaze to enable fine-grained bimanual manipulation with only a single stereo camera. EyeRobot 2.0 physically attends to a 3D fixation point in the scene by swiveling two eye viewpoints to center their gaze on it. The resulting images are processed foveally by allocating more visual tokens to the image centers, focusing computation on task-relevant features. Such Active Visual Fixation (AVF) requires carefully coordinated gaze during task execution, which we accomplish hierarchically by first training a low-level gaze servoing policy conditioned on a goal object, then training a target selector which emits fixation goals based on task progress. Both modules are trained with RL on real-world data: the first is trained with a dense geometric reward and the second co-trains with the BC gripper policy which allows it to discover fixation sequences that can resemble a human's fixation sequence while performing the task. EyeRobot 2.0 further takes advantage of fixation by canonicalizing gripper information into a fixation-relative SE(3) frame, which compacts the size of the action distribution to learn. We collect teleoperation data for 7 real-world and 6 simulated tasks, and conduct over 1000 physical and 1800 simulated robot trials comparing EyeRobot 2.0 against passive stereo and ego + wrist camera policies trained on the same data. Removing wrist cameras is costly for standard policies: with only passive stereo, real-world success drops from 52% to 27%. EyeRobot 2.0 closes this gap with only stereo, outperforming passive stereo by 40% in real and 20% in sim. It matches ego + wrist policies when their wrist views are clear (69% vs. 64%), and more than doubles their success when grasped objects occlude the wrist cameras (48% vs. 22%)

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. To this end, we propose training-time inversion which constrains a frozen video prior by task-bearing source evidence and visible temporal changes, eliminating the need for test-time inversion or candidate ranking. At deployment, the policy generates bimanual and pan/tilt actions-including stay and reacquisition behaviors-from view-aware history, and updates context from newly measured RGB observations. Future-video prediction serves as a co-training signal, while action generation requires neither future-video decoding nor optimal viewpoint annotations. We introduce RoboTwin-AV, a 50-task benchmark with executable pan/tilt control and automatically generated demonstrations. ActiveWAM improves TAVIS out-of-distribution success by up to 17.0 percentage points over the strongest baselines, achieves 20.0 additional points over Fast-WAM on RoboTwin-AV, and outperforms it by 26.7 points on real-world physical kitchen tasks.

Physical active vision allows robots to change their viewpoint when task-relevant observations become unreliable, yet existing manipulation benchmarks provide limited support for studying how policies recover from occlusion during execution. We introduce BAVO-Bench (Bimanual Active Vision under Occlusion), a bimanual active-vision benchmark that systematically controls external visibility through Clean, Stage Occlusion, and Random-time Occlusion conditions, enabling evaluation of both manipulation performance and active visual recovery. Building on this setting, we present A-FAR (Active Future-Aware Recovery), an active-vision policy for joint viewpoint and manipulation control. A-FAR represents moving-camera observations in a unified robot-centric 3D frame and distills relational structure together with its future evolution from a pretrained 4D model, providing the policy with future-aware geometric guidance without requiring future observations at deployment. Experiments across multiple manipulation tasks show that A-FAR improves robustness to both structured and temporally shifted occlusions while maintaining strong performance under clean observations.

Multi-agent manipulation naturally produces multiple task-driven viewpoints: every arm carries a wrist camera and moves through the scene while acting. Yet these observations are typically underutilized, and active perception in manipulation is still often treated as requiring a dedicated sensing agent. We introduce MAAP (Multi-Agent Active Perception), in which every arm is dual-purpose: it executes manipulation actions and, through the wrist camera it carries, simultaneously serves as a moving viewpoint for the team. We pair this with RAIL (Role-Aware Imitation Learning), a controller that predicts each arm's current role alongside its action chunk and conditions action generation on it, representing role-dependent actions within one network. Across four simulated tasks, widening the perception regime lifts average success from 56.5% with a fixed camera to 62.5% with one active wrist view and 70.0% with all of them, while MAAP+RAIL reaches 79.2%. RAIL's additional gain is concentrated on the three-arm Microwave task, where success rises from 47% to 82% on identical multi-wrist inputs. On a dual-arm platform, MAAP+RAIL succeeds in 14 of 20 placement trials compared with 0 of 20 for fixed-view ACT. Collaborative manipulation can thus serve as an active perception mechanism in its own right.

Robots inspecting an assembly must determine which parts are present and whether they are correctly installed. During egocentric assembly assistance, head motion and workpiece handling reveal evidence for these checks, while spoken state confirmations link observations to procedural outcomes. We introduce INSPECT, which learns robot view preferences from records of a smart-glasses assistant that answers part queries and provides next-step guidance. Presence-Invariant TwinSwap (PI-TwinSwap) calibrates object evidence through paired identity interventions. Claim-indexed supervision separates evidence requirements from camera-reproducible observation changes. Object-centered calibration adapts relative view preferences to robot poses, while clause-level screening checks predicted evidence. The robot selects views using only its current observation and known poses, without candidate images. Evaluation uses annotated assistant-video replay to simulate state feedback, without target-domain view labels for policy training. On images of physical gearbox assemblies, INSPECT achieves the highest view utility among the compared non-oracle policies and raises human-rated full verifiability from 34.8% to 41.7% compared with keeping the current view. On commercial angle-grinder recordings in IMPACT, the transferred relative-view selector increases the correct decision rate from 50.6% to 54.3% with a frozen perception head. The source code is available at https://github.com/Kratos-Wen/INSPECT.

Active perception allows autonomous agents to select their viewpoints rather than passively process the viewpoints given to them, enabling them to target where to reduce uncertainty about their environment. Learned systems typically encourage this behavior with hand-designed proxy objectives, such as coverage or curiosity bonuses, that may conflict with the task. In this work, we propose a method to learn emergent active perception (LEAP) without augmentation of the task objective. We formulate the problem of goal-oriented navigation over hazardous terrains with goals that must be discovered visually. We then propose an architecture for navigation policies with active perception, and train them on a terrain curriculum where task pressure alone leads to the emergence of gaze control. Key to this emergence, LEAP works on a gaze-invariant representation that integrates depth images into egocentric belief maps. We validate its performance in held-out evaluation scenarios, where it achieves a 92.7% success rate, compared to 74.2% for scripted or 34.5% for passive perception, and comes within 4.6 points of a privileged oracle. We validate that LEAP navigation policies, unchanged, can be directly applied to steering quadrupedal locomotion policies in physics simulation.

Natural language provides robots with a flexible task interface, but target ambiguity in embodied environments arises not only from user intent; it can also result from missing taskrelevant physical evidence in the current observation. Existing interactive disambiguation methods primarily obtain additional information by asking the user, whereas occlusion, restricted viewpoints, unreadable text, and unobserved targets require the robot to actively change its observation. We propose an active-perception framework for embodied target disambiguation that uses active observation as the backbone for information acquisition and uses a vision-language model to decide, on the basis of accumulated visual evidence and interaction information, whether to continue observing, request clarification, or complete target selection. Active observation can both directly recover missing discriminative evidence and reveal object names, labels, and semantic attributes, thereby improving user clarification when it remains necessary. Real-robot experiments show that the framework combines physical information acquisition and userintent clarification within a unified embodied disambiguation process.

Human vision is a closed loop: gaze is continuously redirected by intermediate hypotheses rather than a single snapshot. Decades of psychophysics and cognitive science have argued that this active observation is essential for a wide range of tasks. Whether today's multimodal large language models (MLLMs) exercise active observation is an empirical question that current vision-language benchmarks do not answer. We introduce ActiveVision, a benchmark that makes active observation measurable for MLLMs, comprising 17 tasks across 3 categories. Tasks are designed to force repeated visual perception rather than a single static description. Frontier MLLMs collapse on ActiveVision: the highest-scoring model we evaluate, GPT-5.5 at the highest exposed reasoning-effort tier, solves only 10.6% of items and scores zero on 11 of the 17 tasks, and even Claude Fable 5, despite topping most reasoning and coding leaderboards, solves just 3.5%, far behind three human participants who average 96.1%. Furthermore, much of the gap persists even when models write and run their own vision code: such code is unreliable on realistic imagery, and catching its failures itself requires the active perception the models lack. Together, these results indicate that current MLLMs lack robust active visual observation, motivating architectures and training objectives that close the perception-reasoning loop.

Autonomous robots often need to move their camera before they can act: to inspect an object, reveal an occluded region, or obtain a view that responds to a user's intent. While vision-language navigation translates instructions to base motion and vision-language-action policies map instructions to manipulation actions, vision-language conditioned camera motion remains comparatively underexplored as a first-class action. We formulate vision-language conditioned camera motion generation: given a current RGB observation and a free-form natural-language intent, predict a relative target camera pose for the next observation. The challenge is that useful viewpoint changes depend on latent perceptual intent, ranging from coarse spatial moves to fine-grained inspection or occlusion revealing. To model this structure, we mine multi-intent camera-motion supervision from egocentric video via hindsight labeling, pairing each relative SE(3) target pose with plausible intents and observation-gain descriptions inferred from the reached view. We propose LIME, a vision-language camera-motion generator that combines an autoregressive observation-gain output with a continuous flow-matching pose head. This design lets the model jointly predict what the next view should reveal while representing multi-hypothesis target views. Across viewpoint-prediction experiments and downstream robotic tasks, we show that LIME learns intent-conditioned camera motion from passive egocentric video, turning ordinary human recordings into supervision for a reusable active-perception primitive that supports general-purpose viewpoint selection, manipulation, embodied question answering, and mid-distance navigation.

Spin plays a crucial role in many ball sports due to its effect on the trajectory of the ball. Vision-based estimation of the ball's spin during a game with conventional cameras is challenging due to the ball's small size, high speed, and fast rotation. To address these challenges, we propose an event-based active vision system that can track unmodified balls and measure their spin in real time. The system consists of an event camera for its high temporal resolution and minimal motion blur, high-speed pan/tilt galvanometer mirrors to keep the ball in the field of view, and a low-latency focus-tunable telephoto lens to increase the spatial resolution on the ball and keep it in focus. To track the ball, we use a hybrid approach that combines 2D event-based detection for centering and 3D positions from a ball localization system for re-initialization. For high-accuracy spin estimation, we propose an offline method that performs contrast maximization on the sphere (s-CMax). This method achieves state-of-the-art accuracy on static balls across multiple sports (table tennis, baseball, tennis, and golf), with mean magnitude and axis errors of 1.2% and 1.5 degrees, respectively. We then develop a low-latency online method for table tennis as a case study in real-time applications. This method uses an uncertainty-aware convolutional neural network trained on pseudo-ground-truth spin labels from the offline approach, combined with a GPU-accelerated batch implementation of contrast maximization for refinement. We demonstrate reliable tracking and spin estimation with a three-view setup during professional table tennis matches, with high accuracy (8.8% magnitude and 6.4 degrees axis mismatch w.r.t. the offline method), 3 ms latency, and 750 Hz throughput.

Agile humanoid locomotion across diverse challenging terrain demands both wide perceptual coverage and precise local geometry understanding. Motivated by the way humans selectively look at relevant terrain during locomotion, we introduce TAGA, a Terrain-aware Active Gaze learning framework for Attention-based humanoid control. By fusing vision, proprioception, and motion commands, our framework guides the model to learn anticipatory cues and actively attend to specific areas of the height scan, selectively using these informative regions for the downstream network. This adaptively increases the information density of observations under tight onboard computational constraints, thus enabling fine-grained perceptive locomotion over larger-scale terrains. We find that such gaze behaviors can naturally emerge through reinforcement learning alone, without requiring additional supervision or explicit guidance, significantly improve training efficiency. As a result, the trained policy demonstrates robust and generalizable locomotion in simulation and on hardware, including reliable terrain-aware foothold selection, elevated-platform traversal, competitive sparse-foothold traversal, and the largest reported real-world gap traversal distance of 1.2m among perceptive humanoid locomotion systems, while maintaining stability under severe perceptual disturbances and environmental interference.

We investigate whether behavior cloning is sufficient to produce active perception in a structured object-finding task. A low-cost robot arm equipped with a wrist-mounted egocentric RGB camera must reposition to center a partially visible plant before triggering a grasp signal, requiring actions that improve future observations. The model predicts joint commands directly from low-resolution RGB images under closed-loop control. We show that low-resolution egocentric vision is sufficient for reliable task completion and that predicting relative joint deltas substantially outperforms absolute joint position prediction in our setting. These results demonstrate that visually grounded active perception can emerge from behavior cloning in a reproducible setting.

Active vision -- where a policy controls its own gaze during manipulation -- has emerged as a key capability for imitation learning, with multiple independent systems demonstrating its benefits in the past year. Yet there is no shared benchmark to compare approaches or quantify what active vision contributes, on which task types, and under what conditions. We introduce TAVIS, evaluation infrastructure for active-vision imitation learning, with two complementary task suites -- TAVIS-Head (5 tasks, global search via pan/tilt necks) and TAVIS-Hands (3 tasks, local occlusion via wrist cameras) -- on two humanoid torso embodiments (GR1T2, Reachy2), built on IsaacLab. TAVIS provides three evaluation primitives: a paired headcam-vs-fixedcam protocol on identical demonstrations; GALT (Gaze-Action Lead Time), a novel metric grounded in cognitive science and HRI that quantifies anticipatory gaze in learned policies; and procedural ID/OOD splits. Baseline experiments with Diffusion Policy and $π_0$ reveal that (i) active vision generally helps, consistently across training runs, but benefits are task-conditional rather than uniform; (ii) multi-task policies degrade sharply under controlled distribution shifts on both suites; and (iii) imitation alone yields anticipatory gaze that lands on the task-relevant object, with lead times comparable to those of the human demonstrations, although head motion is less smooth than in the demonstrations, an effect of action chunking that success rate does not reveal. Code and evaluation scripts are released at https://github.com/spiglerg/tavis; demonstrations (LeRobot v3.0; ~2200 episodes) and trained baselines at https://huggingface.co/tavis-benchmark.

Human visual reasoning is governed by active vision, a process where meta-cognitive control drives top-down goal-directed attention, dynamically routing foveal focus toward task-relevant details while maintaining peripheral awareness of the global scene. In contrast, modern Vision-Language Models (VLMs) process visual information passively, relying on the static accumulation of massive token contexts that dilute the visual evidence as the reasoning chain grows. Here we propose GazeVLM, a VLM that learns to exert this oversight over its attention resources by generating gaze actions in its reasoning chain, in the form of <LOOK> tags specifying the coordinates of the region to inspect. During training, each <LOOK> block triggers a continuous suppression bias on the attention logits that dampens features outside the regions gazed at so far, and that lifts at the end of the block, restoring the global view. The model is trained with this bias as a teaching signal, first by Supervised Fine-Tuning on curated gaze-reasoning traces, then through Group Relative Policy Optimization (GRPO) with rewards for correct answers and valid grounding. At deployment, the external suppression is removed and the model, running its unmodified forward pass, steers its own attention to the regions indicated by the <LOOK> blocks, retaining most of the effect of the bias and emulating top-down control of spatial attention. This model thus transitions between global spatial awareness and localized focal reasoning without relying on external agentic tools like cropping, and without adding visual tokens from re-encoded patches. Applied to 4B-parameter backbones, GazeVLM demonstrates strong high-resolution multimodal reasoning on HRBench-4k and HRBench-8k, surpassing its base models by about 4 points, and outperforming agentic multimodal pipelines built around thinking with images by 5 to 12 points.

Optimizing vision models purely for classification accuracy can impose an alignment tax, degrading human-like scanpaths and limiting interpretability. We introduce EVA, a neuroscience-inspired hard-attention mechanistic testbed that makes the performance-human-likeness trade-off explicit and adjustable. EVA samples a small number of sequential glimpses using a minimal fovea-periphery representation with CNN-based feature extractor and integrates variance control and adaptive gating to stabilize and regulate attention dynamics. EVA is trained with the standard classification objective without gaze supervision. On CIFAR-10 with dense human gaze annotations, EVA improves scanpath alignment under established metrics such as DTW, NSS, while maintaining competitive accuracy. Ablations show that CNN-based feature extraction drives accuracy but suppresses human-likeness, whereas variance control and gating restore human-aligned trajectories with minimal performance loss. We further validate EVA's scalability on ImageNet-100 and evaluate scanpath alignment on COCO-Search18 without COCO-Search18 gaze supervision or finetuning, where EVA yields human-like scanpaths on natural scenes without additional training. Overall, EVA provides a principled framework for trustworthy, human-interpretable active vision.

Active computer vision promises efficient, biologically plausible perception through sequential, localized glimpses, but lacks scalable general-purpose architectures and pretraining pipelines, leaving Active-Vision Foundation Models (AVFMs) underexplored. We introduce CanViT, the first task- and policy-agnostic AVFM. CanViT uses scene-relative RoPE to bind a retinotopic Vision Transformer backbone and a spatiotopic scene-wide latent workspace, the canvas. Efficient interaction with this high-capacity working memory is supported by Canvas Attention, a novel asymmetric cross-attention mechanism. We decouple thinking (backbone-level) and memory (canvas-level), eliminating canvas-side self-attention and fully-connected layers to achieve fast sequential inference and scalability to high output resolutions. We propose a label-free active vision pretraining scheme, policy-agnostic passive-to-active dense latent distillation: reconstructing scene-wide DINOv3 embeddings from sequences of low-resolution glimpses with randomized locations, zoom levels, and lengths. We pretrain CanViT-B from a random initialization on 13.2 million ImageNet-21k scenes--an order of magnitude more than previous active models--and 1 billion random glimpses, in 166 hours on a single H100. On ADE20K segmentation, a frozen CanViT-B achieves 38.5% mIoU in a single low-resolution glimpse, outperforming the best active model's 27.6% with 20x fewer inference FLOPs as well as its FLOP- or input-matched DINOv3 teacher. Given additional glimpses, CanViT-B reaches 45.9% ADE20K mIoU. On ImageNet-1k classification, CanViT-B also sets a new active-vision state of the art, with 84.5% top-1 accuracy after fine-tuning. CanViT generalizes to longer rollouts, larger scenes, and new policies. Our work narrows the wide gap between passive and active computer vision, demonstrating the potential of task- and policy-agnostic AVFM pretraining.

Human vision achieves remarkable perceptual performance while operating under strict metabolic constraints. A key ingredient is the selective attention mechanism, driven by rapid saccadic eye movements that constantly reposition the high-resolution fovea onto task-relevant locations, unlike conventional AI systems that process entire images with equal emphasis. Our work aims to draw inspiration from the human visual system to create smarter, more efficient image processing models. Using DINO, a self-supervised Vision Transformer that produces attention maps strikingly similar to human gaze patterns, we explore a saccade inspired method to focus the processing of information on key regions in visual space. To do so, we use the ImageNet dataset in a standard classification task and measure how each successive saccade affects the model's class scores. This selective-processing strategy preserves most of the full-image classification performance and can even outperform it in certain cases. By benchmarking against established saliency models built for human gaze prediction, we demonstrate that DINO provides superior fixation guidance for selecting informative regions. These findings highlight Vision Transformer attention as a promising basis for biologically inspired active vision and open new directions for efficient, neuromorphic visual processing.

Active perception - the ability of a robot to proactively select viewpoints to acquire task-relevant information - is essential for robust operation in real-world environments. However, existing approaches are typically limited to fixed objectives or constrained settings, and struggle to generalize to open-ended perception intents specified in natural language. We propose I-Perceive, a foundation model for language-conditioned active perception in large-scale indoor environments. Given a query image, a set of context images, and a natural language instruction, I-Perceive predicts a 6D camera pose that fulfills the specified perception intent. The model integrates a vision-language pathway for semantic grounding with a geometric reasoning pathway for multi-view 3D understanding, connected via multi-layer semantic fusion to enable language-conditioned geometric reasoning. To support scalable training, we construct a large-scale dataset of language-viewpoint pairs from both real-world scene-scanning data and simulated environments using an automated pipeline. Extensive experiments demonstrate that I-Perceive significantly outperforms strong baselines on prediction accuracy, viewpoint feasibility, and instructions alignment. The model exhibits strong zero-shot generalization to unseen scenes and instructions, and enables closed-loop active perception, progressively refining viewpoints over sequential interactions.

Active visual perception enables an intelligent system to improve perception by controlling how visual evidence is acquired. However, differences in how active perception is defined and evaluated make approaches difficult to compare across tasks. This review clarifies the field's functional scope by distinguishing active observation acquisition from processing fixed inputs. It then organizes representative literature around four questions about what evidence the task requires, which actions can obtain it, how candidate observations should be valued, and how the next sensing action should be selected and executed. New observations update the perceptual state and guide subsequent decisions, forming a closed loop. To evaluate the benefits of active perception, the review examines datasets, interactive environments, physical platforms, and baseline comparisons, and develops an evaluation framework linking improvements in visual evidence to downstream task outcomes while accounting for acquisition costs and risks. The analysis concludes that predicted observation value does not guarantee that useful evidence can be acquired through feasible actions, while task success alone cannot establish the benefit of active observation selection. Further progress requires reliable value estimation and planning over longer decision sequences under physical and resource constraints. It also requires grounding semantic reasoning in acquired evidence and maintaining effective collaboration under imperfect communication, with these capabilities verified through reproducible evaluation of the complete perception and action loop.