Pushing the Limits of Reactive Planning: Learning to Escape Local Minima

TL;DR

The paper addresses the challenge of navigating cluttered environments without relying on a full map by augmenting a fast, purely reactive planner with neural components that inject geometric intuition. It combines Riemannian Motion Policies and ray-based sensing as a safety layer with neural networks (FFN and LSTM) trained in synthetic worlds, using a privileged geodesic distance field during training to shape decisions without requiring maps at run-time. The key contributions include multiple neural-reactive architectures, a detailed analysis of how memory influences navigation, and a demonstration of zero-shot transfer to real 3D environments, including resilience to sensor noise. The work advances practical reactive navigation by closing the loop between classical local policies and learned geometric priors, enabling safer and more capable obstacle avoidance and goal pursuit in complex spaces.

Abstract

When does a robot planner need a map? Reactive methods that use only the robot's current sensor data and local information are fast and flexible, but prone to getting stuck in local minima. Is there a middle-ground between fully reactive methods and map-based path planners? In this paper, we investigate feed forward and recurrent networks to augment a purely reactive sensor-based planner, which should give the robot geometric intuition about how to escape local minima. We train on a large number of extremely cluttered worlds auto-generated from primitive shapes, and show that our system zero-shot transfers to real 3D man-made environments, and can handle up to 30% sensor noise without degeneration of performance. We also offer a discussion of what role network memory plays in our final system, and what insights can be drawn about the nature of reactive vs. map-based navigation.

Figures (7)

• Figure 1: A comparison of baseline (blue) and learned LSTM (rainbow) trajectories from various starting points to the goal location shown by a green sphere. The reactive baseline cannot navigate to the goal when it is directly behind a wall, while our learned method uses its "geometric intuition" (shown by the high influence of the LSTM) to turn left to escape the local minimum. The world is from the BundleFusion dataset dai2017bundlefusion, and the learned system is transferred zero-shot to this environment after training on synthetically generated data containing only obstacle primitives.
• Figure 2: System overview of the learned reactive planner. We combine the local avoidance policies as a safety layer with a higher level learned system, supervised by the geodesic field during training.
• Figure 3: Example of 1024 ray-casts obtained via Halton sampling. These rays are a map-agnostic method to gain local geometric information.
• Figure 4: Example of 2 purely reactive planning runs getting stuck in the same local minimum. The goal location is behind the wall, depicted in green. The world is from the BundleFusion dataset dai2017bundlefusion.
• Figure 5: Network architecture. Learned blocks are highlighted in green. All learned blocks consist of a linear layer followed by layer normalization and a Leaky ReLU, except for the final decoder layer, which contains only a single linear layer. Layer normalization is used over alternatives like batch normalization, as it is more compatible with sequenced data. The LSTM is only included in the recurrent neural reactive navigation system.