Learning Enabled Fast Planning and Control in Dynamic Environments with Intermittent Information

Matthew Cleaveland, Esen Yel, Yiannis Kantaros, Insup Lee, Nicola Bezzo

TL;DR

This paper introduces a novel framework for safe robot navigation in dynamic environments where sensor data is only available intermittently. By training a neural network offline to approximate computationally expensive reachability-based potential fields, the authors enable real-time, safety-critical planning that accounts for obstacle motion uncertainty even when operating on "stale" information.

What Problem Does This Paper Address?

Imagine an autonomous underwater vehicle (UUV) trying to cross a busy shipping lane. It can only receive information about the positions and headings of surface ships when it briefly surfaces. Once it dives, it is "blind" to the dynamic obstacles around it, relying on data that becomes increasingly outdated (stale) as time passes.

Standard motion planning often assumes continuous, high-frequency sensor feedback. In communication- or sensor-limited environments—like underwater, underground, or in adversarial military zones—this assumption fails. The challenge is twofold:

  1. Uncertainty: The robot must predict where obstacles might be, accounting for their noisy dynamics and the time elapsed since the last update.
  2. Computation: Reachability analysis (RA) is the gold standard for calculating these "safety sets," but it is far too slow to run in a real-time control loop.

Methods

The authors propose a two-stage compositional neural network (NN) controller framework:

  1. Offline Stage: The researchers use Reachability Analysis (e.g., using the Flow tool) to compute the set of all possible future states an obstacle could occupy given its last known position, heading, and bounded noise. They then define a repulsive potential field* based on these reachable sets. To solve the speed issue, they train an NN to take obstacle state and "time since last detection" as inputs and output the gradient of this repulsive field.
  2. Online Stage: At runtime, the robot uses the pre-trained NN to quickly generate avoidance headings for every known obstacle. Because the gradient operator is linear, the system is compositional: the robot simply sums the avoidance vectors from the NN for any number of obstacles and adds a "Go-to-Goal" (GTG) vector.
Architecture of the proposed approach, showing the split between offline training and online execution.

Datasets

The authors did not use a standard public dataset. Instead, they generated massive synthetic datasets for training their NNs based on the specific dynamics of their test vehicles:

  • UUV Case Study: ~8 million data points generated using obstacle dynamics and Flow* reachability analysis.
  • UGV Experiments: ~3 million data points. The training data maps robot/obstacle relative states and time intervals to the required repulsive force gradients.

Key Insights and Ideas

  • Baking Safety into Weights: The primary innovation is "baking" the expensive reachability calculations into the weights of a neural network. This transforms a minute-long calculation into a millisecond-scale inference.
  • Solving Dynamic Minima: Traditional potential fields often suffer from "dynamic minima," where a robot gets stuck directly in front of a moving obstacle because the repulsive force pushes it back toward its starting point. By using the temporal properties of reachability sets (which are often conic/wedge-shaped), the repulsive force naturally pushes the robot behind the obstacle's predicted path.
  • Compositionality: By training the NN to handle a single obstacle, the framework scales effortlessly. Whether there are two ships or twenty, the robot just runs the NN for each and sums the results.
Example of a dynamic minima in standard potential field planning vs. the proposed method.

Key Results

The framework was validated through both high-fidelity ROS Gazebo simulations and real-world lab experiments with Clearpath Jackal and TurtleBot2 UGVs.

  • Speed: In the UUV simulation, calculating a reachable set 600 seconds into the future took an average of 67 seconds using standard tools. The NN controller achieved the same result in an average of 27ms (max 81ms), enabling real-time control.
  • Safety: In 20 trials of a UUV crossing a channel with ships moving at 5m/s, the robot safely navigated using information that was up to 9 minutes old. Planners that ignored motion uncertainty resulted in collisions in 55% of the trials.
  • Real-world Robustness: UGV experiments confirmed that the robot could successfully avoid two dynamic obstacles moving in opposite or identical directions, even after losing all sensor contact at the start of the mission.
UUV crossing a shipping channel. The blue squares represent the reachable sets (uncertainty) of the ships.

Limitations and Future Work

While the results are promising, the authors note several areas for improvement:

  • Localization Uncertainty: The current model assumes the robot knows its own position perfectly; future work will incorporate "ego-vehicle" uncertainty.
  • Cooperative Agents: The model currently treats all obstacles as non-cooperative. Integrating cooperative multi-robot logic could improve efficiency.
  • Formal Verification: While the reachability analysis provides safety guarantees, the NN is an approximation. The authors plan to investigate verification tools to provide formal safety bounds for the NN component itself.

Summary updated

Abstract

This paper addresses a safe planning and control problem for mobile robots operating in communication- and sensor-limited dynamic environments. In this case the robots cannot sense the objects around them and must instead rely on intermittent, external information about the environment, as e.g., in underwater applications. The challenge in this case is that the robots must plan using only this stale data, while accounting for any noise in the data or uncertainty in the environment. To address this challenge we propose a compositional technique which leverages neural networks to quickly plan and control a robot through crowded and dynamic environments using only intermittent information. Specifically, our tool uses reachability analysis and potential fields to train a neural network that is capable of generating safe control actions. We demonstrate our technique both in simulation with an underwater vehicle crossing a crowded shipping channel and with real experiments with ground vehicles in communication- and sensor-limited environments.
BibTeX
@inproceedings{cleaveland2022learning,
  author = {Matthew Cleaveland and Esen Yel and Yiannis Kantaros and Insup Lee and Nicola Bezzo},
  title = {{Learning Enabled Fast Planning and Control in Dynamic Environments with Intermittent Information}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2022},
  doi = {10.1109/iros47612.2022.9981508},
}
Figure