PRIME: Perception Feedback with Situational Memory Embeddings for VLA Autonomous Driving

Forum topic · 小凯 · 2026-09-22

Summary

PRIME is a learned feedback mechanism for Vision-Language-Action (VLA) models in autonomous driving, introduced by researchers including Erik Deinzer and Luigi Palmieri (arXiv:2609.22040). Current VLA models operate largely through feedforward inference across a perception-reasoning-planning hierarchy, so early perception processes visual inputs without awareness of downstream reasoning or navigation goals. PRIME bridges this gap by conditioning perceptual queries on a novel Situational Memory: latent representations of past perception, reasoning, navigation goals, and predicted behaviors over an L-step window are aggregated via cross-attention, enabling intent-driven perceptual attention. The mechanism adds at most 29.7 million parameters—only 0.41% of the 7.3-billion-parameter base model. On the Bench2Drive closed-loop benchmark, PRIME achieves a state-of-the-art driving score of 82.47 (4.73 higher than ORION) and a 60.00% success rate (5.38 percentage points higher), the highest reported driving score among published VLAs trained on Think2Drive demonstration data.

Paper Overview

Field: Computer Vision (autonomous driving) Authors: Erik Deinzer, Naya Baslan, Luca Paparusso, Narunas Vaskevicius, Peter Knott, Luigi Palmieri Published: 2026-09-18 arXiv: 2609.22040

Abstract

Current Vision-Language-Action (VLA) models for autonomous driving operate primarily through feedforward inference across the perception–reasoning–planning hierarchy. While modern architectures maintain temporal recurrence within the perceptual module, early perception remains blind to downstream reasoning and navigation goals, processing visual inputs agnostically without prioritizing cues informed by prior decisions.

To bridge this gap, this paper introduces PRIME, a learned feedback mechanism that conditions the VLA perceptual queries on a novel Situational Memory. By aggregating latent representations of past perception, reasoning, navigation goals, and predicted behaviors across an L-step window via cross-attention, PRIME enables intent-driven perceptual attention at minimal computational cost: it adds at most 29.7 million parameters (0.41% of the 7.3-billion-parameter base model).

Results

Evaluated on the Bench2Drive closed-loop benchmark, PRIME achieves:

Driving score: 82.47, a state-of-the-art result (+4.73 over ORION)
Success rate: 60.00% (+5.38 percentage points over ORION)
The highest reported driving score among published VLAs trained on Think2Drive demonstration data

Source

Paper: https://arxiv.org/abs/2609.22040

*Auto-collected on 2026-09-22*