Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference

2026-05-04

AI Summary

Existing dual-system vision-language-action (VLA) models suffer from inefficiency due to the computationally expensive per-step invocation of the vision-language model (VLM) backbone and high redundancy among features from adjacent frames. This work proposes Latent Bridge, a lightweight, architecture-agnostic mechanism that predicts the difference between VLM features across consecutive timesteps, enabling the action head to operate on predicted features while invoking the VLM backbone only periodically. The approach supports task-agnostic DAgger training and seamlessly transfers across multiple benchmarks without architectural modifications. By integrating feature-space bridging (GR00T-N1.6) and KV-cache bridging (π0.5), Latent Bridge achieves 95–100% performance retention on LIBERO, RoboCasa, and ALOHA benchmarks while reducing VLM invocations by 50–75% and accelerating inference by 1.65–1.73×.

Abstract

Dual-system Vision-Language-Action (VLA) models achieve state-of-the-art robotic manipulation but are bottlenecked by the VLM backbone, which must execute at every control step while producing temporally redundant features. We propose Latent Bridge, a lightweight model that predicts VLM output deltas between timesteps, enabling the action head to operate on predicted outputs while the expensive VLM backbone is called only periodically. We instantiate Latent Bridge on two architecturally distinct VLAs: GR00T-N1.6 (feature-space bridge) and π0.5 (KV-cache bridge), demonstrating that the approach generalizes across VLA designs. Our task-agnostic DAgger training pipeline transfers across benchmarks without modification. Across four LIBERO suites, 24 RoboCasa kitchen tasks, and the ALOHA sim transfer-cube task, Latent Bridge achieves 95-100% performance retention while reducing VLM calls by 50-75%, yielding 1.65-1.73x net per-episode speedup.

Problem

• Vision-Language-Action
• VLM backbone
• temporal redundancy
• efficient inference
• robotic manipulation

Innovation

• Latent Bridge
• Vision-Language-Action (VLA)
• feature delta prediction
• efficient inference
• KV-cache