Neural Autonomous Navigation with Riemannian Motion Policy

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

TL;DR

This paper introduces a novel image-based autonomous navigation framework that combines deep learning with Riemannian Motion Policies (RMPs). By training a neural network to predict RMPs rather than direct control commands or depth maps, the system explicitly models vehicle geometry and dynamics, resulting in a highly interpretable, generalizable controller capable of navigating tight spaces and avoiding obstacles both in simulation and the real world.

What Problem Does This Paper Address?

Autonomous navigation using cheap, lightweight cameras is highly desirable, but traditional SLAM (Simultaneous Localization and Mapping) pipelines struggle in textureless environments or with fast-moving cameras. To bypass SLAM, recent deep learning approaches typically fall into two camps:

  1. Predicting local geometry (like depth maps) from images, which are then fed into traditional planners. However, depth estimation from a single image can be ill-posed (e.g., when facing a blank wall), leading to unpredictable and catastrophic control failures.
  2. Predicting control commands directly (end-to-end learning). These networks act as black boxes. Because they don't explicitly model the robot's geometry or physical dynamics, they struggle to generalize to unseen environments and offer no way to interpret why the robot made a specific decision.

The authors tackle this by asking: How can we build a vision-based neural controller that retains the explicit geometric and dynamic modeling of classical robotics, while leveraging the pattern-matching power of deep learning?

Methods

The authors bridge the gap between vision-based learning and model-based control using the Riemannian Motion Policy (RMP) framework.

  • RMP Basics: An RMP models the interaction between a point on the robot and the environment using an acceleration policy paired with a Riemannian metric (which dictates the "cost" or importance of moving in certain directions).
  • Vehicle Modeling: They model an Ackermann steering vehicle (an RC car) using 12 distinct "control points" distributed around its bounding box. Goal RMPs pull the front points toward a target, while Obstacle RMPs push points away from hazards.
  • Neural RMP: Instead of calculating RMPs from perfect laser scans, they train a ResNet-50-based neural network to predict the RMPs (accelerations and metrics for each control point) directly from a monocular RGB image and basic odometry (velocity, goal location).
  • Analytical Solver: The predicted RMPs are fed into a deterministic, non-neural solver. This solver incorporates the vehicle's kinematic constraints (e.g., a car can't move sideways) to analytically compute the optimal steering and acceleration commands.
  • Baselines: They compare their Neural RMP against two structurally similar networks: one that predicts a 1D depth scan, and one that predicts control commands directly.

Datasets

  • Training: The Gibson simulation environment. The dataset consists of 60,000 trajectories sampled from 15 indoor spaces. Ground truth was generated by an "expert" RMP controller using simulated 240-degree laser scans.
  • Data Augmentation: DAGGER (Dataset Aggregation) was used during training to expose the network to recovery behaviors.
  • Testing: 5 holdout Gibson environments (space8, house24, house29, house31, house57) and a real-world physical hallway.
  • Real-world Finetuning: ~13,000 real-world images paired with laser scans were collected to finetune the models and bridge the sim-to-real domain gap.

Key Insights and Ideas

  • Geometry + Dynamics = Generalization: By forcing the network to output RMPs rather than raw steering angles, the system inherently respects the physical footprint and non-holonomic constraints of the car. This prevents the network from suggesting physically impossible maneuvers, leading to vastly better generalization in unseen environments.
  • Interpretability via Metrics: Unlike black-box end-to-end models, the Neural RMP is interpretable. You can visualize the predicted Riemannian metrics as elliptical isocontours around the car's control points. High costs are visibly assigned to directions facing obstacles, explaining exactly why the car decides to turn or brake.
  • Robustness over Depth Prediction: Predicting RMPs is more concise and less noisy than predicting dense depth maps. In tight spaces where a small depth error would normally cause a crash, the RMP representation remains stable.

Key Results

  • Simulation Superiority: In the 5 holdout Gibson environments, the Neural RMP consistently outperformed the baselines. For example, in "house29", the RMP agent reached the goal 93.7% of the time with only a 1.6% collision rate. The direct control baseline only reached the goal 68.5% of the time (14.2% collision rate).
  • Complex Maneuvers: In tight spaces requiring sharp turns, the direct control baseline often crashed into walls. The RMP agent successfully learned to back up to create headroom for sharp turns, mimicking the expert behavior.
  • Real-World Transfer: After a brief 20-minute finetuning on real hallway data, the Neural RMP successfully navigated a physical hallway at 25 frames per second. In a tricky scenario where the car started off facing a wall, the direct control agent crashed straight into it, while the RMP agent intelligently backed up and navigated around the obstacle.

Limitations and Future Work

  • Limitations: The system suffers from a noticeable sim-to-real domain gap; models trained purely in Gibson failed in the real hallway without manual finetuning. Additionally, because the RMP is fundamentally a local reactive policy, it can get stuck in local minima (e.g., trapped behind a large, concave obstacle) if the provided waypoint is directly behind the obstacle.
  • Future Work: The authors suggest pairing the Neural RMP with a high-level neural planner to provide better global waypoints. They also propose exploring unsupervised RMP learning, applying the framework to agile robot maneuvers, and extending it to 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},
}

Loading related content...

Figure