Reproduce “A Large Recurrent Action Model: xLSTM enables Fast Inference for Robotics Tasks” 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
Thomas Schmied, Thomas Adler, Vihang Patil, Maximilian Beck, Korbinian Pöppel, Johannes Brandstetter, Günter Klambauer, Razvan Pascanu and 1 more
A Large Recurrent Action Model (LRAM) based on xLSTM and Mamba architectures demonstrates high-performance, real-time inference for robotics tasks, overcoming the computational limitations of Transformer-based models. It achieves competitive or superior performance on a diverse suite of 432 tasks while offering significantly faster inference speeds, crucial for real-time control.
Key insights
- Modern recurrent architectures like xLSTM and Mamba offer linear-time inference complexity, making them inherently more suitable than Transformers for real-time applications with long sequence lengths.
- Pre-training on diverse, large-scale multi-domain datasets allows recurrent models to learn transferable representations that enhance performance on new tasks through fine-tuning.
- xLSTM's unique sLSTM blocks contribute to superior state-tracking capabilities and strong in-context learning performance, especially in partially observable environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.