Reproduce “Kinaema: a recurrent sequence model for memory and pose in motion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mert Bulent Sariyildiz, Philippe Weinzaepfel, Guillaume Bono, Gianluca Monaci, Christian Wolf
Kinaema gives robots a recurrent memory they can query with a destination image to estimate relative pose. It encodes camera images and odometry, then updates distributed memory embeddings using transformer attention and gating. In simulated tests, it reaches 28% pose-recognition accuracy on 800-step sequences and 54% SPL for hard, primer-seen goals; the study does not establish real-world performance.
Key insights
- Combining transformer-based contextualization with efficient recurrence and a learned gating mechanism is effective for building scalable and stable long-term spatial memory in embodied agents.
- Training Kinaema with randomized sequence lengths and an auxiliary Masked Image Modeling (MIM) loss is crucial for robust generalization to significantly longer, out-of-distribution observation histories.
- Kinaema's latent memory tokens learn stable, spatially coherent correspondences to observed scene regions over time, demonstrating the emergence of an implicit, spatially aware internal model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.