BIT-Nav: Brain-Inspired Trajectory Memory for Embodied Navigation

2026-06-19

AI Summary

Existing vision-language models (VLMs) for embodied navigation rely on sparse frame sampling, which struggles to capture structured behaviors—such as turning patterns, displacement accumulation, and path topology—over long-horizon trajectories. Inspired by hippocampal path integration mechanisms, this work proposes a learnable, compact trajectory memory module that encodes sequences of actions and relative poses using a bidirectional GRU, compressing motion histories of arbitrary length into a fixed-dimensional memory embedding. This embedding is injected as a single token into a frozen VLM. By combining a multi-positive InfoNCE contrastive loss with a lightweight MLP projection, the method enables long-horizon trajectory understanding without increasing inference overhead, substantially enhancing the model’s capacity to reason about long-term behavioral intent and overcoming the performance limitations imposed by conventional sparse sampling strategies.

Abstract

Vision-Language Models (VLMs) for embodied navigation rely on selecting a fixed number of frames from a growing trajectory history. As episodes extend, this selection grows increasingly sparse, yet prior work shows no accuracy gain when scaling from 8 to 64 frames, suggesting the bottleneck is not frame quantity but the representation itself. Sparse frame selection cannot capture the structured behavioral signal that long-horizon reasoning requires: turning patterns, cumulative displacement, and path topology. We introduce BIT-Nav (Brain-Inspired Trajectory Memory for Navigation), a framework that augments frozen VLM navigation pipelines with a compact learned trajectory memory. Motivated by hippocampal path integration, where spatial experience is compressed into structured episodic traces rather than stored as raw sensory replay, BIT-Nav trains a Bi-GRU encoder over action and relative pose sequences via a multi-positive InfoNCE contrastive objective on trajectory prefixes sharing the same behavioral intent. The resulting embedding is projected into the VLM token space via a lightweight MLP and injected as a single memory token at each decision step, conditioning the model on structured motion history at constant token cost regardless of episode length

Problem

• embodied navigation
• trajectory memory
• structured behavioral signal
• long-horizon reasoning
• vision-language models

Innovation

• trajectory memory
• brain-inspired navigation
• contrastive learning
• vision-language models
• path integration