Neural Autonomous Navigation with Riemannian Motion Policy

Xiangyun Meng, Nathan D. Ratliff, Yu Xiang, Dieter Fox

Abstract

End-to-end learning for autonomous navigation has received substantial attention recently as a promising method for reducing modeling error. However, its data complexity, especially around generalization to unseen environments, is high. We introduce a novel image-based autonomous navigation technique that leverages in policy structure using the Riemannian Motion Policy (RMP) framework for deep learning of vehicular control. We design a deep neural network to predict control point RMPs of the vehicle from visual images, from which the optimal control commands can be computed analytically. We show that our network trained in the Gibson environment can be used for indoor obstacle avoidance and navigation on a real RC car, and our RMP representation generalizes better to unseen environments than predicting local geometry or predicting control commands directly.

Introduction

Autonomous navigation using low-cost, monocular cameras is a significant challenge in robotics. Traditional SLAM-based pipelines often fail in textureless environments or during rapid movement. To address this, deep learning approaches have emerged, typically falling into two categories: predicting local geometry (e.g., depth maps) for traditional planners, or end-to-end learning that predicts control commands directly.

However, both have drawbacks. Depth estimation can be ill-posed in featureless areas, leading to catastrophic failures. Direct control prediction often results in "black-box" models that lack interpretability and struggle to generalize because they do not explicitly model the robot's physical footprint or dynamics. This paper proposes a middle ground: a vision-based neural controller that retains explicit geometric and dynamic modeling through the Riemannian Motion Policy (RMP) framework.

Overview of the Neural RMP navigation pipeline and its application to an RC car with designated control points.

Methods

The core of this approach is the integration of the RMP framework with deep learning. An RMP models the interaction between a robot's control points and its environment using an acceleration policy paired with a Riemannian metric, which represents the "cost" of moving in specific directions.

Vehicle Modeling and Neural RMP

The authors model an Ackermann steering vehicle using 12 control points distributed around its bounding box. Instead of calculating RMPs from sensor data like laser scans, a ResNet-50-based neural network is trained to predict these RMPs—specifically the accelerations and metrics for each control point—directly from a monocular RGB image and basic odometry (velocity and goal location).

Analytical Control

The predicted RMPs are processed by a deterministic, non-neural solver. This solver incorporates the vehicle's kinematic constraints analytically to compute optimal steering and acceleration commands. This ensures that the output is physically feasible for the vehicle.

Architectural comparison between (a) the proposed Neural RMP predictor and baselines for (b) depth prediction and (c) direct control prediction.

Experimental Setup

The system was trained in the Gibson simulation environment using 60,000 trajectories from 15 indoor spaces. Ground truth was provided by an "expert" RMP controller utilizing simulated laser scans. To improve robustness, the DAGGER (Dataset Aggregation) algorithm was employed to expose the network to recovery behaviors.

The Neural RMP was compared against two baselines:

  1. Predicting Depth: A network predicting a 1D depth scan.
  2. Predicting Controls: A network predicting control commands directly.

Testing was conducted across five holdout Gibson environments and a real-world physical hallway. For the real-world tests, the models were finetuned using approximately 13,000 real-world images to bridge the sim-to-real gap.

Key Results

  • Superior Generalization: In holdout simulation environments, the Neural RMP significantly outperformed baselines. In one environment ("house29"), it achieved a 93.7% success rate with only a 1.6% collision rate, compared to the direct control baseline's 68.5% success and 14.2% collision rate.
  • Complex Maneuverability: The RMP agent successfully learned complex behaviors, such as backing up to navigate sharp turns in tight spaces—a behavior the direct control baseline failed to replicate reliably.
  • Interpretability: The predicted Riemannian metrics can be visualized as elliptical isocontours, providing a clear explanation of the car's decisions by showing high costs assigned to directions containing obstacles.
  • Real-World Performance: After 20 minutes of finetuning, the Neural RMP navigated a real hallway at 25 FPS. It demonstrated superior obstacle avoidance compared to the direct control agent, which crashed when faced with a wall.

Conclusion and Future Work

The Neural RMP framework successfully combines the pattern-matching strengths of deep learning with the structural benefits of classical model-based control. While it shows a domain gap between simulation and reality requiring finetuning, it offers a more interpretable and generalizable solution than standard end-to-end models. Future work includes integrating high-level neural planners to avoid local minima and extending the framework to more complex robotic manipulation tasks.

Summary updated

Abstract

End-to-end learning for autonomous navigation has received substantial attention recently as a promising method for reducing modeling error. However, its data complexity, especially around generalization to unseen environments, is high. We introduce a novel image-based autonomous navigation technique that leverages in policy structure using the Riemannian Motion Policy (RMP) framework for deep learning of vehicular control. We design a deep neural network to predict control point RMPs of the vehicle from visual images, from which the optimal control commands can be computed analytically. We show that our network trained in the Gibson environment can be used for indoor obstacle avoidance and navigation on a real RC car, and our RMP representation generalizes better to unseen environments than predicting local geometry or predicting control commands directly.
BibTeX
@inproceedings{meng2019neural,
  author = {Xiangyun Meng and Nathan D. Ratliff and Yu Xiang and Dieter Fox},
  title = {{Neural Autonomous Navigation with Riemannian Motion Policy}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2019},
  url = {https://openreview.net/forum?id=jlrNx9E9eg},
}
Figure