Kinaema: a recurrent sequence model for memory and pose in motion

TL;DR

Kinaema introduces a distributed latent memory updated by a transformer to enable continual localization of previously observed spaces without storing full observation histories, achieving $O(1)$ update/read complexity per step. It defines Mem-RPE, a relative-pose task between a query image and memory, and Mem-Nav, a continuous navigation setting with a priming phase before goal exposure, and demonstrates integration into a downstream navigation agent for improved performance. The approach yields state-of-the-art Mem-RPE results against strong baselines and significantly boosts Mem-Nav SPL when memory-augmented with Kinaema, illustrating practical gains for long-horizon embodied AI. The work highlights memory scalability, stable training via gating, and cross-attention-based decoding as key enablers for memory-aware navigation with potential impact on real-world continual robotic operation.

Abstract

One key aspect of spatially aware robots is the ability to "find their bearings", ie. to correctly situate themselves in previously seen spaces. In this work, we focus on this particular scenario of continuous robotics operations, where information observed before an actual episode start is exploited to optimize efficiency. We introduce a new model, Kinaema, and agent, capable of integrating a stream of visual observations while moving in a potentially large scene, and upon request, processing a query image and predicting the relative position of the shown space with respect to its current position. Our model does not explicitly store an observation history, therefore does not have hard constraints on context length. It maintains an implicit latent memory, which is updated by a transformer in a recurrent way, compressing the history of sensor readings into a compact representation. We evaluate the impact of this model in a new downstream task we call "Mem-Nav". We show that our large-capacity recurrent model maintains a useful representation of the scene, navigates to goals observed before the actual episode start, and is computationally efficient, in particular compared to classical transformers with attention over an observation history.

Figures (15)

• Figure 1: We introduce Kinaema, a model capable of situating previously observed spaces: a recurrent transformer compresses observed sequences into latent memory and estimates rel. pose of a goal image w.r.t. to its current state.
• Figure 2: Kinaema is a recurrent sequence model maintaining distributed memory$\mathbf{m}_t$ in the form of $N$ embeddings of size $E$ each. Its previous state $\mathbf{m}_{t-1}$ is first contextualized with observations $\mathbf{o}_t=\{\textbf{x}_t, \mathbf{u}_t\}$ then embedding-wise gated, resulting in new state $\mathbf{m}_t$.
• Figure 3: Integration into the downstream RL-trained agent: The RPE decoder used for pre-training is kept for the downstream task, searching for the goal image in the embeddings $\mathbf{y}_t$ which are read out from memory $\mathbf{m}_t$, while a binocular encoder $BinEnc$ from DEBiTCrocoNav2024 is used to compare the goal to the current observation. Two types of memory are updated and queried: (i) the agent maintains its own additional recurrent memory $\mathbf{h}_t$, and (ii) Kinaema-memory $\mathbf{m}_t$.
• Figure 4: Generalization to longer sequences, Mem-RPE: GRU and Kinaema, trained for T=100, evaluated on $T{=}100...1000$ (RPE-test).
• Figure 5: RPE during navigation: exploiting the information from the priming sequence (green), Kinaema can predict (pink) the rel. pose of the goal (red), while the binocular module of CrocoNav2024 only starts providing reliable predictions (orange) when the agent is positioned (cyan) in view of the goal.