IROS 2026 · ACCEPTED
Observation-Conditioned Rollout Allocation for Sampling Model Predictive Control on Myopic Egocentric Elevation Maps
IEEE/RSJ International Conference on Intelligent Robots and Systems · IROS 2026
Reactive local navigation for a car-like UGV using observation-conditioned predictive sampling and myopic egocentric 2.5D terrain information.
Research overview Publications
Overview
Sampling-based Model Predictive Control evaluates many candidate motions during each control cycle.
When the robot only has access to a limited egocentric representation of its surroundings, distributing a fixed rollout budget uniformly can spend valuable samples in directions that are not especially useful for the current terrain observation.
This work studies how the available rollout budget can instead be allocated according to what the robot currently observes.
The proposed method concentrates predictive sampling in locally relevant regions while preserving explicit trajectory evaluation and model-based control.
Real-world demonstration
The method was evaluated on a real outdoor car-like unmanned ground vehicle using onboard perception and computation.
Your browser does not support embedded video.
:::
The video shows the navigation system operating on the real platform during an outdoor field experiment.
Navigation problem
The robot operates with a limited sensor horizon and therefore does not have access to a complete map of the environment.
Navigation decisions must be made from a myopic egocentric 2.5D terrain representation while satisfying the motion constraints of a car-like vehicle and the real-time requirements of the predictive controller.
The main challenge addressed in this work is therefore not to increase the total number of predictive samples, but to use a fixed sampling budget more effectively.
Proposed approach
The method modifies how the available MPPI rollouts are distributed around the robot.
1 · Observe
The robot builds an egocentric 2.5D representation of the terrain currently visible to its sensors.
2 · Allocate
The fixed rollout budget is distributed across structured local regions according to the current observation.
3 · Evaluate
Candidate trajectories are propagated and evaluated by the predictive controller before the final vehicle command is selected.
The observation-conditioned component influences the sampling process. It does not directly command the robot.
Observation-conditioned rollout allocation
The local sampling space is divided into structured regions around the vehicle.
Rather than assigning the same amount of sampling effort everywhere, the allocation changes according to the terrain currently observed by the robot.
Conceptually, the navigation loop is:
Egocentric terrain observation
↓
Local region analysis
↓
Observation-conditioned rollout allocation
↓
MPPI candidate trajectories
↓
Trajectory evaluation
↓
Vehicle commandThe total rollout budget remains bounded, allowing the method to remain suitable for real-time execution.
Evaluation
The experimental study compares observation-conditioned rollout allocation with uniform sampling under the same general predictive-control setting.
The evaluation considers navigation performance together with real-time execution, including metrics related to success, safety events, trajectory behaviour and computational latency.
Main contribution
The central contribution is a method for reallocating a fixed MPPI sampling budget according to the current egocentric terrain observation.
This allows the predictive controller to spend more of its available computation on locally relevant motion hypotheses without replacing the explicit model-based control process.
Relation to my PhD
This paper represents one stage of my current doctoral research on autonomous UGV navigation in unstructured terrain.
It focuses on reactive local navigation using only the terrain currently visible around the robot.
My ongoing research extends this direction toward terrain-aware local control, navigation using global DEM information, autonomous exploration and compact terrain memory.
See the complete research overview
Material
Additional public material will be linked here as it becomes available:
Paper
IEEE Xplore page
IROS 2026 poster
Conference presentation
Demonstration video
Supplementary material
Public code, where applicable