ReViP: Reducing False Completion in Vision-Language-Action Models with Vision-Proprioception Rebalance

TL;DR

Vision-Language-Action models suffer false completion when proprioceptive progress dominates visual evidence. ReViP addresses this by introducing Vision–Proprioception Rebalance with a Task-Stage Observer (external VLM) and a Task-Stage Enhancer (TS-FiLM) to inject task-centric visual cues into the VLA backbone, improving visual grounding and robustness under perturbations. It also provides the False-Completion Benchmark Suite to systematically evaluate this failure mode. Across simulation benchmarks (LIBERO, RoboTwin 2.0) and real-world tests, ReViP reduces false completions and boosts success, and demonstrates plug-and-play compatibility across backbones, advancing robust, visually grounded manipulation.

Abstract

Vision-Language-Action (VLA) models have advanced robotic manipulation by combining vision, language, and proprioception to predict actions. However, previous methods fuse proprioceptive signals directly with VLM-encoded vision-language features, resulting in state-dominant bias and false completions despite visible execution failures. We attribute this to modality imbalance, where policies over-rely on internal state while underusing visual evidence. To address this, we present ReViP, a novel VLA framework with Vision-Proprioception Rebalance to enhance visual grounding and robustness under perturbations. The key insight is to introduce auxiliary task-aware environment priors to adaptively modulate the coupling between semantic perception and proprioceptive dynamics. Specifically, we use an external VLM as a task-stage observer to extract real-time task-centric visual cues from visual observations, which drive a Vision-Proprioception Feature-wise Linear Modulation to enhance environmental awareness and reduce state-driven errors. Moreover, to evaluate false completion, we propose the first False-Completion Benchmark Suite built on LIBERO with controlled settings such as Object-Drop. Extensive experiments show that ReViP effectively reduces false-completion rates and improves success rates over strong VLA baselines on our suite, with gains extending to LIBERO, RoboTwin 2.0, and real-world evaluations.

Figures (7)

• Figure 1: Rebalancing Vision & Proprioception. Under unexpected perturbations such as object drops, existing VLA models may exhibit false completion by prioritizing internal state progression over visual feedback. By injecting task-stage cues, ReViP rebalances semantic perception and proprioceptive dynamics, enabling the policy to identify the object, re-pick it, and complete the task, leading to consistent performance improvements across simulation benchmarks and real-world experiments.
• Figure 2: Overview of the proposed ReViP framework. It consists of two stages: (above) Task-Stage Observer for cues extraction and (below) Task-Stage Enhancer for rebalancing. Through a frozen vision–language model (Qwen 2.5), task–centric visual cues are extracted from the observation $O_t$ and task instructions. These cues are then injected into the VLA backbone by the Task-Stage Enhancer to adaptively rebalance the visual and proprioceptive streams.
• Figure 3: Illustrative examples of our False Completion Benchmark. The benchmark consists of three complementary perturbation sources. Object drops assess whether the policy can detect failures that occur during execution. Distractor swaps examine instance-level grounding under visually similar objects. Relayout conditions test spatial reasoning when the target and its goal appear in new configurations.
• Figure 4: Qualitative comparisons on our False Completion benchmark with Object Drop settings (above: simulation, below: real-world). For simulation, ReViP detects the object drop during execution, and successfully re-picks the target (salad dressing), achieving a true completion $\smiley$, while $\pi_0$ fails to react to the clear visual failure and continues executing a state-dominant bias, resulting in false completion $\frownie$. In the real world, ReViP can also retrieve the toy when it falls, but $\pi_0$ still fails to respond correctly.
• Figure 5: The experimental platform consists of a robot arm, gripper, and cameras. The task objects include multiple targets such as toys and cubes, covering both deformable and rigid items. Data collection and model inference are conducted using RGB-D cameras from both first-person and third-person viewpoints.