Risk-Averse RRT* Planning with Nonlinear Steering and Tracking Controllers for Nonlinear Robotic Systems Under Uncertainty

Sleiman Safaoui, Benjamin J. Gravell, Venkatraman Renganathan, Tyler H. Summers

TL;DR

This paper introduces RANS-RRT*, a novel motion planning algorithm for nonlinear robotic systems operating under uncertainty. It combines a risk-aware planner with robust low-level controllers, demonstrating significantly improved safety and performance in cluttered environments compared to existing methods.

What Problem Does This Paper Address?

Deploying robots safely in uncertain and dynamic environments, like urban streets or crowded airspaces, is a major challenge. Existing motion planning algorithms often assume deterministic or linearly-behaved systems, which don't reflect the reality of nonlinear robot dynamics and unpredictable disturbances. This can lead to unsafe trajectories that might collide with obstacles, especially when dealing with heavy-tailed noise. The paper tackles this by developing a planning framework that explicitly accounts for these uncertainties and risks, aiming to generate trajectories that are not only feasible but also demonstrably safe.

Methods

The core contribution is the Risk-Averse Nonlinear Steering RRT (RANS-RRT) algorithm. This is a two-phase approach:

  1. High-Level Planning (RANS-RRT*):
    • Nonlinear Steering: Instead of simple geometric steering, RANS-RRT uses a Nonlinear Program (NLP) to generate dynamically feasible trajectory segments between nodes in the RRT tree. This NLP incorporates the robot's nonlinear dynamics.
  2. Distributionally Robust (DR) Collision Checking: To account for uncertainty, RANS-RRT* approximates the state distribution using the Unscented Transformation (UT) and performs collision checks based on a moment-based ambiguity set. This "risk-tightened" collision check ensures that even under worst-case distributions within the estimated uncertainty, the robot avoids obstacles with high probability.
  3. RRT* Framework: The algorithm builds upon the RRT* framework to ensure asymptotic optimality, meaning it will find the optimal path if given enough time.
  4. Low-Level Tracking:
    • The trajectory generated by RANS-RRT* serves as a reference for a low-level controller.
  5. Three controllers are evaluated:
  6. Linear Quadratic Regulator (LQR): A standard controller that linearizes dynamics around the reference.
  7. Robust LQR (LQRm): An extension of LQR that incorporates robustness-promoting multiplicative noise terms to handle model uncertainties.
  8. Nonlinear Model Predictive Control (NMPC): A more advanced controller that directly optimizes for the nonlinear dynamics and constraints, offering better tracking performance.

Datasets

Not discussed in the paper. The experiments use simulated unicycle dynamics in custom-designed cluttered environments.

Key Insights and Ideas

  • DR Collision Checking is Crucial for Safety: The paper strongly emphasizes that simply assuming Gaussian noise or using standard collision checks is insufficient for nonlinear systems. The distributionally robust approach, which considers a broader range of possible noise distributions, is key to achieving reliable safety guarantees.
  • High-Level Risk Awareness Pays Off: Even with relatively small assumed noise levels in the high-level planner, the resulting trajectories are significantly safer. This highlights the importance of incorporating risk awareness early in the planning process.
  • NMPC Outperforms LQR Variants in High Noise: While LQR and robust LQR are simpler, NMPC demonstrates superior performance in tracking reference trajectories under significant process noise, achieving fewer collisions and better state deviation control.
  • RANS-RRT* Generates Conservative but Safe Paths: The algorithm naturally avoids areas close to obstacles, leading to trajectories that might seem overly cautious but are essential for safe operation in uncertain environments.

Key Results

  • RANS-RRT* Trees: The algorithm successfully generated feasible and safe RRT trees in complex environments, avoiding risky gaps that a standard RRT might exploit.
  • Tracking Performance:
    • At low noise levels ($\sigma^2 = 0.0000005$), all controllers performed well, with minimal differences.
  • At higher noise levels ($\sigma^2 = 0.0035$), RANS-RRT*-generated trajectories with NMPC tracking showed significantly fewer collisions and better reference tracking compared to LQR and robust LQR.
  • For instance, at $\sigma^2 = 0.0035$, NMPC had 75 collisions out of 1000 trials, while LQR had 160 and robust LQR had 138. Open-loop control failed almost entirely (999 collisions).
  • Computational Cost: NMPC incurred a higher computational cost (longer run time) due to its more complex optimization problem.

Limitations and Future Work

  • Approximations in Risk Treatment: The paper acknowledges that the risk treatment step is an approximation, and the covariance propagation using UT is imperfect, meaning formal risk-bound guarantees are not provided.
  • Simplified Sensor Model: The current work assumes perfect state estimation. Future work should incorporate a full nonlinear sensor model.
  • Computational Efficiency: The authors aim to decrease NMPC computation time using code generation tools.
  • Exact DR Risk Constraints: Future research will focus on incorporating exact DR risk constraints into the optimization problems at both planning and control levels for more accurate risk assessment.

Summary updated

Abstract

We propose a two-phase risk-averse architecture for controlling stochastic nonlinear robotic systems. We present Risk-Averse Nonlinear Steering RRT* (RANS-RRT*) as an RRT* variant that incorporates nonlinear dynamics by solving a nonlinear program (NLP) and accounts for risk by approximating the state distribution and performing a distributionally robust (DR) collision check to promote safe planning. The generated plan is used as a reference for a low-level tracking controller. We demonstrate three controllers: finite horizon linear quadratic regulator (LQR) with linearized dynamics around the reference trajectory, LQR with robustness-promoting multiplicative noise terms, and a nonlinear model predictive control law (NMPC). We demonstrate the effectiveness of our algorithm using unicycle dynamics under heavy-tailed Laplace process noise in a cluttered environment.
BibTeX
@inproceedings{safaoui2021risk,
  author = {Sleiman Safaoui and Benjamin J. Gravell and Venkatraman Renganathan and Tyler H. Summers},
  title = {{Risk-Averse RRT* Planning with Nonlinear Steering and Tracking Controllers for Nonlinear Robotic Systems Under Uncertainty}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2021},
  url = {https://openreview.net/forum?id=hdN8iu5VLg},
}
Figure