Abstract
Long-range navigation is commonly addressed through hierarchical pipelines in which a global planner generates a path, decomposed into waypoints, and followed sequentially by a local planner. These systems are sensitive to global path quality, as inaccurate remote sensing data can result in locally infeasible waypoints, which degrade local execution. At the same time, the limited global context available to the local planner hinders long-range efficiency. To address this issue, we propose a reinforcement learning-based local navigation policy that leverages path information as contextual guidance. The policy is conditioned on reference path observations and trained with a reward function mainly based on goal-reaching objectives, without any explicit path-following reward. Through this implicit conditioning, the policy learns to opportunistically exploit path information while remaining robust to misleading or degraded guidance. Experimental results show that the proposed approach significantly improves navigation efficiency when high-quality paths are available and maintains baseline-level performance when path observations are severely degraded or even non-existent. These properties make the method particularly well-suited for long-range navigation scenarios in which high-level plans are approximate and local execution must remain adaptive to uncertainty.
Introduction
Most long-range autonomous navigation systems rely on a hierarchical structure: a global planner analyzes a map to find a path, and a local planner executes immediate movements to follow that path. However, this structure is inherently brittle. If the global map is outdated or low-resolution, the global planner might suggest a path through a new obstacle or impassable terrain. Traditional local planners often try to follow these "bad" waypoints blindly, leading to collisions or stuck states. Conversely, if the local planner ignores the global context entirely, it lacks the "big picture" and can easily get trapped in dead ends.
Researchers at ETH Zurich have developed a reinforcement learning (RL) local planner that treats global paths as "helpful suggestions" rather than strict commands. By training without an explicit path-following reward, the robot learns to opportunistically exploit global guidance when it's useful but ignore it when it's noisy, suboptimal, or leads into a wall.
Methodology
The core of the proposed system is a path-conditioned RL framework. Unlike standard planners, this model observes a sequence of 15 waypoints simultaneously.
Key innovations in the architecture include:
- Path Encoding Module: Uses Self-Attention to understand the global structure of the path (like upcoming turns) and Cross-Attention with a learnable query to let the policy focus on the most relevant segments for the current situation.
- Implicit Conditioning: The reward function is based almost entirely on goal-reaching and safety, rather than path adherence. The robot only follows the path if doing so actually helps it reach the goal faster or more safely.
- Shortcut Reward: An explicit reward encourages the robot to make more progress toward the goal than the suggested path would have allowed, incentivizing the discovery of local shortcuts.
- Robust Training: During simulation, the agent is fed "degraded" paths—intentionally noisy or locally infeasible—forcing it to learn when to trust guidance and when to rely on its own depth cameras.
Experimental Setup
The model was trained and evaluated using Isaac Sim / Isaac Lab, utilizing 180 procedurally generated maze-like environments. While training occurred on 30x30m maps, generalization was tested on larger 50x50m maps.
To validate the approach in the real world, the system was deployed on a Unitree B2W quadruped robot. Equipped with a ZEDX depth camera and an NVIDIA Jetson AGX Orin, the robot successfully navigated a university building over trajectories exceeding 90 meters.
Results and Discussion
The proposed model demonstrates a significant boost in efficiency and robustness. When provided with an optimal global path, it achieved a 7.02% increase in Success weighted by Path Length (SPL) compared to a baseline RL planner that only received a single waypoint.
Crucially, the model exhibits "graceful degradation." When fed "bad" paths (noisy and 3x longer than optimal), it maintained a success rate of 0.83, matching the baseline performance. It did not allow poor advice to degrade its performance below that of a standard "blind" explorer. In physical tests, the robot successfully navigated complex indoor environments, choosing safer paths (such as avoiding stairs) even when the global path was imperfect, demonstrating successful sim-to-real transfer.
Conclusion and Future Work
By treating global plans as heuristics rather than constraints, this RL-based local planner achieves a balance between long-range guidance and local adaptivity. However, some limitations remain. The current model uses a fixed window of 15 waypoints, which may limit look-ahead distance for kilometer-long trajectories. Additionally, the robot's trust in the path is implicit; future iterations could benefit from explicit signaling when the robot chooses to ignore a path, which would aid in human-robot interaction and system debugging.
Summary updated