Risk-aware MPPI for Stochastic Hybrid Systems

Hardik Parwana, Mitchell Black, Bardh Hoxha, Hideki Okamoto, Georgios Fainekos, Danil Prokhorov, Dimitra Panagou

TL;DR

This paper introduces a risk-aware variant of Model Predictive Path Integral (MPPI) control designed for stochastic hybrid systems, such as robots navigating around humans with attention zones. By leveraging the Expansion-Compression Unscented Transform (ECUT), the framework maps state uncertainty directly to the stochasticity of mode-switching boundaries, enabling real-time, safety-critical path planning without relying on computationally expensive Monte Carlo simulations.

What Problem Does This Paper Address?

Navigating a mobile robot through a crowd of dynamic agents (like humans) is challenging because agent behaviors change abruptly based on context. For example, a pedestrian might switch from an "uncooperative" mode (ignoring the robot) to a "cooperative" mode (actively avoiding the robot) once the robot enters their field of view or attention zone. Mathematically, this is modeled as a stochastic hybrid system, where discrete mode transitions are triggered by continuous states subject to noise.

Existing robotic path planners face two major hurdles here:

  1. Deterministic Switching Pitfalls: Most current planners decide when a mode switch occurs based solely on the mean of the predicted state distribution. This ignores the reality that the switch itself is probabilistic due to state uncertainty.
  2. Computational Intractability: While Monte Carlo (MC) simulations can capture these complex probabilistic interactions, they require a massive number of samples, making them too slow for real-time, receding-horizon control on actual robots.

Methods

The authors propose ECUT-MPPI, a gradient-free, sampling-based model predictive control framework that integrates the Expansion-Compression Unscented Transform (ECUT) into the MPPI pipeline.

Here is how the novel components fit together:

  • Sigma-Point Mode Switching: Instead of evaluating the mode-switching function using only the mean predicted state, ECUT-MPPI evaluates the switching condition for every individual sigma point (representative particles of the uncertainty distribution). This captures the probabilistic nature of crossing a boundary (e.g., entering a human's field of view), as shown in the conceptual diagram below.
Illustration of robot trajectory inducing mode switches across human sigma points at $t=6$.
  • Scalable Uncertainty Propagation: Passing sigma points through stochastic dynamics normally causes the number of points to explode exponentially over time. The authors use ECUT's moment-matching compression step to collapse the expanded points back down to a fixed number at each time step, keeping the algorithm computationally efficient.
  • Risk-Aware Cost Function: The standard MPPI objective is modified to penalize paths based on a central confidence interval bound ($\mu_{h,p} - \alpha \sigma_{h,p}$), where $\mu$ and $\sigma$ are the mean and standard deviation of the distance to obstacles, and $\alpha$ is a user-defined risk-tolerance parameter.

Datasets

The method was evaluated using simulated environments rather than static datasets:

  1. Multi-Agent Simulation (JAX): A custom decentralized simulation featuring a mobile robot, 2 circular obstacles, and 10 dynamic agents moving via a stochastic potential field model.
  2. AWS Hospital Environment (Gazebo): A 3D simulator benchmark representing a realistic hospital corridor where humans are simulated using a complex Social Force Model (SFM) to test the algorithm's robustness against model mismatch.

Key Insights and Ideas

  • The Switch is Probabilistic: The core "aha" moment is that state uncertainty turns deterministic spatial boundaries (like a field of view) into fuzzy, probabilistic transitions. Accounting for this at the particle level prevents the planner from making overconfident, unsafe assumptions near transition zones.
  • Exploiting Human Awareness Pays Off: If a robot knows that a human will notice it and step aside, it can plan a tighter, more efficient path. However, it must balance this optimism with risk-awareness in case the human fails to notice it due to positioning uncertainty.
  • GPU Acceleration for Sampling: Implementing the MPPI pipeline in JAX allows massive parallelization of the trajectory rollouts, making the unscented transform approach viable for real-time deployment.

Key Results

The proposed ECUT-MPPI was compared against a "dynamics-unaware" baseline (assuming humans never cooperate) and a "mean-based" baseline (switching modes based only on the mean state).

  • Safety and Efficiency: The dynamics-aware ECUT-MPPI achieved the lowest cumulative cost while maintaining zero safety constraint violations within a 95% confidence interval.
  • Baseline Failures: Both the mean-based and dynamics-unaware baselines suffered from safety violations, with the minimum distance to humans and obstacles dropping below zero over time.
Fig. 2: Simulation results for proposed method in Section V-A. (a) The robot performs collision avoidance with humans (red) and obstacles (black). Robot trajectory is similarly shown in green. (b) Cost (15a) and stability cost (Qc only) accumulated with time. (c), (d) Robot’s minimum distance to any (c) human and (d) obstacle with time. The solid/dotted lines and shaded areas represent the mean and the 95% confidence interval in variation over 50 runs.
  • Real-Time Performance: In the Gazebo hospital environment, by limiting trajectory predictions to the nearest 4 humans, the JAX-based controller successfully ran at 20 Hz on an NVIDIA RTX 4060 Laptop GPU using 10,000 samples.

Limitations and Future Work

  • Dimensionality Scaling: The state space grows rapidly with the number of agents. Even with software optimizations, tracking all 10 agents simultaneously in real-time remains a bottleneck, requiring the authors to heuristically limit tracking to the nearest 4 agents in Gazebo.
  • Known Model Structure: The framework assumes the robot has prior knowledge of the underlying structure of the hybrid dynamics modes (e.g., knowing the potential field equations for the human), even if the parameters are stochastic.
  • Next Steps: The authors plan to validate the framework on physical robotic hardware and explore its performance across different classes of hybrid dynamical systems.

Summary updated

Abstract

Path Planning for stochastic hybrid systems presents a unique challenge of predicting distributions of future states subject to a state-dependent dynamics switching function. In this work, we propose a variant of Model Predictive Path Integral Control (MPPI) to plan kinodynamic paths for such systems. Monte Carlo may be inaccurate when few samples are chosen to predict future states under state-dependent disturbances. We employ recently proposed Unscented Transform-based methods to capture stochasticity in the states as well as the state-dependent switching surfaces. This is in contrast to previous works that perform switching based only on the mean of predicted states. We focus our motion planning application on the navigation of a mobile robot in the presence of dynamically moving agents whose responses are based on sensor-constrained attention zones. We evaluate our framework on a simulated mobile robot and show faster convergence to a goal without collisions when the robot exploits the hybrid human dynamics versus when it does not.
BibTeX
@inproceedings{parwana2025risk,
  author = {Hardik Parwana and Mitchell Black and Bardh Hoxha and Hideki Okamoto and Georgios Fainekos and Danil Prokhorov and Dimitra Panagou},
  title = {{Risk-aware MPPI for Stochastic Hybrid Systems}},
  booktitle = {ACC},
  year = {2025},
  doi = {10.23919/acc63710.2025.11107447},
}
Figure