MPPI-Based Dynamic Object Tracking
Abstract
This project implements Model Predictive Path Integral (MPPI) control for dynamic object tracking. A vision pipeline detects and tracks moving targets, while MPPI samples control rollouts in real time to generate collision-free trajectories for a mobile robot. The system runs at control rates suitable for hardware deployment and remains robust to sensor noise and partial occlusion.
Technical Implementation
Control Pipeline
- Perception: Object detection and centroid tracking from RGB stream.
- Prediction: Short-horizon target motion model to estimate future poses.
- Sampling: MPPI rollouts with control noise to evaluate feasible trajectories.
- Cost Design: Target error, smoothness
- Execution: Real-time control updates applied to the robot base.
Stack
- Python / C++ bindings for MPPI core
- ROS for messaging and transforms
- OpenCV for detection and visual tracking
- GPU-accelerated sampling (optional) for higher control rates
Key Features
Real-Time MPPI
Runs at control rates suitable for hardware, with adjustable rollout counts for CPU/GPU targets.
Collision-Aware
Penalizes proximity to static and dynamic obstacles, keeping trajectories safe in cluttered scenes.
Robust Tracking
Handles noisy detections and short occlusions using filtered target state and motion priors.
Results and Impact
The controller maintains < 10 cm steady-state tracking error in simulation and remains stable during 1–2 s occlusion events. Sampling budgets scale gracefully, enabling deployment on embedded GPUs for field robots.
Conclusion
MPPI provides a flexible, sampling-based alternative to classical MPC for fast-moving targets. The modular perception-to-control stack can be adapted to aerial, ground, or manipulator platforms with minimal retuning.