ActiveScale: Scaling Active Perception for Robots across Model, Data, and Hardware

Active perception in robotic manipulation is crucial for addressing the issue of incomplete information from fixed viewpoints. Recent advances in vision-language-action models (VLAs) have significantly broadened robots' ability to perform diverse tasks across environments. However, as tasks become more complex, observations from a fixed viewpoint are often only partially informative. Successfully completing a task may require the robot to actively acquire task-relevant information in cluttered environments, for example, by reorienting its egocentric view to inspect a bowl hidden inside a cabinet or the contents of a backpack, thereby reducing occlusion.

Active perception tasks require a model to infer its camera pose and predict subsequent camera motion, with the latter contingent on the former. However, existing VLAs lack these capabilities. Although egocentric data naturally contains camera motion, existing methods do not effectively exploit this signal for learning active perception. Current approaches predominantly study active perception in static environments, leaving a substantial gap to real-world mobile manipulation, where both the robot and its observation viewpoint must adapt to dynamic, cluttered settings.

The ActiveScale framework advances active perception through coordinated model, data, and hardware designs. The model augments VLA with historical video observations and explicit camera-pose supervision, using per-frame pose tokens and a lightweight prediction head to associate observations across viewpoints and support coherent scene understanding. Data-wise, a scalable human-robot mid-training recipe using 1000 hours of egocentric and robotic data adapts the model to temporal inputs and pose supervision. Hardware-wise, the Active-perception Mobile-manipulation Platform (AMP) supports single-operator teleoperation.

• �� Model Design: Extend input from a single image to a video sequence containing historical frames. Introduce learnable camera tokens to predict translation, rotation, and field of view. • Data and Mid-training: Mid-train on large-scale egocentric and robot data, totaling over 1000 hours. • Hardware Platform: Develop AMP integrating bimanual manipulation, independently actuated camera, and mobile-base control. A single operator can jointly control viewpoint, manipulation, and base motion, providing a scalable interface.

Experimental design includes five real-world tasks covering container and articulated occlusion, below-table perception, multi-stage viewpoint transitions, and large vertical workspace variation. Each task uses 150 demonstrations for post-training and evaluates each method over 20 real-world rollouts under matched conditions. A rollout succeeds only if all required stages and the task-specific final-state criterion are satisfied. We report strict task success rate (SR) and task progress (TP).

ActiveScale achieves higher success rates across all five task families, increasing mean SR by 30%, outperforming baseline models. Human-robot co-training significantly enhances model adaptability in complex environments, increasing success rates by 20%. Ablation studies validate the contributions of camera-pose-aware modeling and egocentric mid-training, with superior performance across five tasks.

ActiveScale framework can be applied to robotic manipulation tasks in dynamic, complex environments. Suitable for scenarios requiring active acquisition of task-relevant information, such as inspecting hidden items or adjusting viewpoints to reduce occlusion. Industry impact includes improved efficiency and accuracy of automated operations.

The model may fail in complex occlusion scenarios, requiring further optimization. High hardware platform costs limit widespread application. Dataset diversity needs enhancement to improve model generalization.