ProactiveVLA: Augmenting Embodied Memory through Proactive Environment Exploration

Shizuo Tian, Haodong Luo, Yutong Li, Yuebing Song, Yunxin Liu, Yuanchun Li · N/A · 2026

Summary

Rapid adaptation to a new environment requires a robot to acquire useful knowledge about local objects, states, and interactions from limited experience.

Abstract Summary

Rapid adaptation to a new environment requires a robot to acquire useful knowledge about local objects, states, and interactions from limited experience. Systems that combine a reasoning agent with a frozen vision-language-action model (VLA) can adapt through execution feedback and memory, making the choice of experience central to their effectiveness. Repeated practice of a target task may refine a familiar solution while leaving other interactions relevant to changed conditions untested. We introduce ProactiveVLA, which uses proactive environment exploration to acquire reusable knowledge for deployment-time adaptation. After completing an initial task, the agent allocates the remaining interaction budget to self-proposed goals covering object affordances, state-changing interactions, and compositions of interactions. It verifies execution outcomes and consolidates both task-directed and exploratory experience into memory that guides subsequent planning and control. ProactiveVLA outperforms the baselines under the same turn budget on LIBERO-Pro and RoboCasa365 Composite-Seen. On LIBERO-Pro Goal-T, with at most one VLA primitive invocation allowed during evaluation, ProactiveVLA completes 48% of instances, compared with 19% for the state-of-the-art task-refinement baseline.

Key Points

Rapid adaptation to a new environment requires a robot to acquire useful knowledge about local objects, states, and interactions from limited experience.
Systems that combine a reasoning agent with a frozen vision-language-action model (VLA) can adapt through execution feedback and memory, making the choice of experience central to their effectiveness.
Repeated practice of a target task may refine a familiar solution while leaving other interactions relevant to changed conditions untested.
We introduce ProactiveVLA, which uses proactive environment exploration to acquire reusable knowledge for deployment-time adaptation.
After completing an initial task, the agent allocates the remaining interaction budget to self-proposed goals covering object affordances, state-changing interactions, and compositions of interactions.

Methodology

ProactiveVLA supports rapid adaptation by acquiring environment experience, consolidating it into a global memory, and using that memory to guide subsequent behavior. A high-level agent interacts with the environment through a set of robot primitives, such as moving to a target pose, adjusting the robot posture, controlling the gripper, and VLAs.

Each adaptation instance consists of a scene and an environment-provided task, with a shared budget of $B$ interaction turns for task execution and autonomous exploration.

Methodology — ProactiveVLA: Augmenting Embodied Memory through Proactive Environment Exploration — Figure 1: Overview of ProactiveVLA. After completing the environment-provided task, the agent uses the remaining interaction budget to explore scene-related goals involving object affordances, state-changing interactions, and interaction compositions. An independent visual checker assesses exploratory outcomes and supplies feedback for continued exploration. Memory management preserves the complete interaction records, organizes experience within scenes, synthesizes knowledge across scenes, and independently reviews candidate rules. Accepted rules form a frozen global memory that guides execution across tasks and variants.

Experiments

We evaluate ProactiveVLA with one reference instance per task, where the agent is given a fixed interaction budget before being evaluated, testing whether proactively acquired environment knowledge transfers beyond the conditions encountered during adaptation. Harness VLA serves as the task-refinement baseline. Both ProactiveVLA and Harness VLA receive the same reference instance.

Analysis

Figure 2 reports the fraction of evaluation instances completed with at most $k$ VLA primitive invocations. Figure 2 (a) reports our Goal-T re-evaluation for the VLA-invocation analysis. ProactiveVLA completes 48% of instances with at most one VLA call, compared with 19% for Harness VLA.

The strongest separation therefore appears at the smallest call budget, where additional controller calls are unavailable to compensate for a less effective execution strategy.

Analysis — ProactiveVLA: Augmenting Embodied Memory through Proactive Environment Exploration — Figure 3: Execution under target occlusion. Both agents move the arm after the target becomes occluded. ProactiveVLA re-localizes before resuming the task and succeeds, while Harness VLA skips re-localization, acts at the wrong position, and fails.

Sources and Demonstrations

Figures are reproduced from Tian et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.