HumanMPC - Safe and Efficient MAV Navigation among Humans

Simon Schaefer, Helen Oleynikova, Sandra Hirche, Stefan LeuteneggerarXiv 2025 · DOI · arXiv

Abstract

Navigating Micro Air Vehicles (MAVs) in human-populated environments requires a delicate balance between safety and efficiency. Traditional methods often fall into the "freezing robot" problem, where overly conservative safety constraints paralyze the robot in complex scenarios. This paper presents HumanMPC (also referred to as HumanHalo), a novel Model Predictive Control framework that utilizes minimally conservative safety constraints and 3D human motion forecasting to enable agile, safe navigation in full 3D space.

The Challenge of 3D Human-Robot Interaction

Navigating a drone around people is a high-stakes balancing act. If a robot is too aggressive, it risks collisions; if it is too conservative—a common issue with formal safety methods like Hamilton-Jacobi reachability—it stops moving entirely because it cannot find a path that is 100% guaranteed to be safe for all possible future human actions.

Furthermore, most existing solutions simplify the problem by treating humans as 2D points. However, humans are complex 3D obstacles with moving limbs. To operate truly safely, a robot must reason about full-body 3D motion in real-time on limited onboard hardware, a task that is typically computationally prohibitive for traditional safety-critical controllers.

The HumanMPC Framework

HumanMPC addresses these challenges by combining data-driven human motion forecasting with a novel reachability-based safety formulation. The framework is built on three core pillars:

1. Minimally Conservative MPC

The core innovation is a "twist" on reachability. Instead of forcing the entire planned trajectory to be safe, the controller only ensures that the initial control input does not lead to a state where a collision becomes inevitable. As long as there is at least one sequence of future actions that keeps the robot safe, the current action is allowed. This maintains recursive feasibility without paralyzing the robot.

2. Hybrid Human Reachable Sets

To maintain real-time performance, the system uses a two-tiered model for the human "danger zone":

  • Short-term (Complex): For the immediate future (up to 0.2s), the human is modeled as a collection of 14 capsules and spheres following the skeleton's joints.
  • Long-term (Simplified): For the remainder of the planning horizon, the human is modeled as a single expanding cylinder.

3. Efficient Zonotope Representation

The MAV’s reachable set is modeled as a zonotope—a computationally efficient geometric shape that allows safety constraints to be expressed as linear inequalities. This transforms the complex navigation problem into a Quadratic Program (QP) that can be solved in milliseconds on embedded hardware.

System architecture of HumanMPC, illustrating the perception stack, reachable set construction for both the MAV and human, and the final constraint evaluation for safe control.

Key Insights

  • Recursive Safety via Initial Input: By ensuring the first step always leaves an "out," the robot maintains safety while preserving agility.
  • Intentional Asymmetry: The framework uses an exact model for the MAV but a conservative "worst-case" over-approximation for the human, ensuring safety even during sudden, unpredictable human movements.
  • Linearization for Speed: By linearizing MAV dynamics and using zonotopes, the system avoids heavy non-linear optimizations, enabling high-frequency (40Hz) control loops.

Experimental Results

The system was validated using the AMASS dataset for diverse human motion simulations and deployed on a physical quadrotor equipped with an Intel RealSense D455 and a Jetson Orin NX.

Experimental setup showing a quadrotor autonomously navigating in close proximity to a human using the HumanMPC framework.
  • Safety: In simulations with multiple humans, HumanMPC achieved a 100% collision avoidance rate, outperforming baselines like ORCA (80-88%).
  • Efficiency: HumanMPC maintained high success rates in reaching goals while keeping "Time-to-Goal" significantly lower than other safe controllers.
  • Real-Time Performance: Solver times on the Jetson Orin NX ranged from 3.9ms to 8.4ms, well within the requirements for real-time flight.

Limitations and Future Work

While HumanMPC provides a robust framework for human-robot interaction, it currently relies on a clear line of sight for perception. Future work will focus on handling occlusions and integrating these dynamic constraints with maps of static obstacles like furniture and walls. Expanding the framework to include other dynamic non-human obstacles is also a key next step.

Summary updated

BibTeX
@misc{schaefer2025humanmpc,
  author = {Simon Schaefer and Helen Oleynikova and Sandra Hirche and Stefan Leutenegger},
  title = {{HumanMPC - Safe and Efficient MAV Navigation among Humans}},
  year = {2025},
  eprint = {2510.17525},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.17525},
}
Figure