Despite their remarkable success in language modeling, transformers trained to predict the next token in a sequence struggle with long-term planning. This limitation is particularly evident in tasks requiring foresight to plan multiple steps ahead such as maze navigation. The standard next single token prediction objective, however, offers no explicit mechanism to predict multiple steps ahead - or revisit the path taken so far. Consequently, in this work we study whether explicitly predicting multiple steps ahead (and backwards) can improve transformers' maze navigation. We train parameter-matched transformers from scratch, under identical settings, to navigate mazes of varying types and sizes with standard next token prediction and MLM-U, an objective explicitly predicting multiple steps ahead and backwards. We find that MLM-U considerably improves transformers' ability to navigate mazes compared to standard next token prediction across maze types and complexities. We also find MLM-U training is 4x more sample efficient and converges 2x faster in terms of GPU training hours relative to next token training. Finally, for more complex mazes we find MLM-U benefits from scaling to larger transformers. Remarkably, we find transformers trained with MLM-U outperform larger transformers trained with next token prediction using additional supervision from A* search traces. We hope these findings underscore the promise of learning objectives to advance transformers' capacity for long-term planning. The code can be found at https://github.com/facebookresearch/maze_navigation_MLMU

Lightweight autonomous unmanned aerial vehicles (UAV) are emerging as a central component of a broad range of applications. However, autonomous navigation necessitates the implementation of perception algorithms, often deep neural networks (DNN), that process the input of sensor observations, such as that from cameras and LiDARs, for control logic. The complexity of such algorithms clashes with the severe constraints of these devices in terms of computing power, energy, memory, and execution time. In this paper, we propose NaviSplit, the first instance of a lightweight navigation framework embedding a distributed and dynamic multi-branched neural model. At its core is a DNN split at a compression point, resulting in two model parts: (1) the head model, that is executed at the vehicle, which partially processes and compacts perception from sensors; and (2) the tail model, that is executed at an interconnected compute-capable device, which processes the remainder of the compacted perception and infers navigation commands. Different from prior work, the NaviSplit framework includes a neural gate that dynamically selects a specific head model to minimize channel usage while efficiently supporting the navigation network. In our implementation, the perception model extracts a 2D depth map from a monocular RGB image captured by the drone using the robust simulator Microsoft AirSim. Our results demonstrate that the NaviSplit depth model achieves an extraction accuracy of 72-81% while transmitting an extremely small amount of data (1.2-18 KB) to the edge server. When using the neural gate, as utilized by NaviSplit, we obtain a slightly higher navigation accuracy as compared to a larger static network by 0.3% while significantly reducing the data rate by 95%. To the best of our knowledge, this is the first exemplar of dynamic multi-branched model based on split DNNs for autonomous navigation.