Trajectory-Level Analysis of Neural Trajectory Prediction Models for Smart Wheelchair Navigation
Author: Pouya Bathaei Pourmand Affiliation: MSc in Computer Engineering (AI) — University of Genoa / CNR, Italy Project: REXASI-PRO — Reliable & Explainable AI for Smart Mobility (EU Horizon Europe) Supervisors: Prof. Luca Oneto · Prof. Maurizio Mongelli · Dr. Sara Narteni
This repository contains the public results and source code from the MSc thesis:
Analysis of Neural Trajectory Prediction for Collision Avoidance in Smart Wheelchairs
The thesis develops a systematic trajectory-level evaluation framework for pretrained neural motion prediction models (DNN-LNA) used in autonomous wheelchair navigation.
The central research question is:
When and why do neural trajectory predictors fail?
All five models are treated as black-box predictors and are evaluated exclusively through their outputs, without access to internal weights, training procedures, or retraining.
⚠️ The original DNN-LNA model weights are not included because they are proprietary assets of the REXASI-PRO project. Only derived results, figures, tables and evaluation code are publicly available.
- Systematic trajectory-level evaluation of five DNN-LNA models
- Input-space risk mapping using sampled navigation commands
- Goal-based performance analysis using strict and soft success metrics
- Explainability through Decision Tree models
- Identification of critical failure regions for smart wheelchair navigation
- Comparative analysis of model robustness under different operating conditions
A reproducible demonstration is available in:
notebooks/demo_analysis.ipynb
The notebook reproduces the main methodology using synthetic data that follows the operational ranges of the REXASI-PRO wheelchair platform. No proprietary model weights or confidential datasets are required.
- Experiment 1 — Input-space sensitivity analysis
- Experiment 2 — Goal-based performance analysis
- Decision Tree explainability
- Risk maps and failure region visualization
- Summary of key findings
Note: The notebook uses synthetic data exclusively for demonstration purposes. Real experimental results are provided in the repository under the
results/directory.
Each navigation scenario is defined by a target goal position and three motion commands:
| Symbol | Variable | Operational Range |
|---|---|---|
| GoalX | Goal X coordinate | Goal dependent |
| GoalY | Goal Y coordinate | Goal dependent |
| θ | Initial orientation | [−π, π] rad |
| v | Linear velocity | [−1.05, 2.88] m/s |
| ω | Angular velocity | [−1.99, 1.99] rad/s |
These inputs are provided to a pretrained DNN-LNA model, which predicts a future trajectory:
X̂ = f(GoalX, GoalY, θ, v, ω)
The output trajectory consists of a sequence of future wheelchair states over a prediction horizon of 30 time steps.
Trajectory quality is evaluated using two criteria:
d < 0.30 m
- Success: d < 0.30 m
- Near-success: 0.30 ≤ d < 0.50 m
- Failure: d ≥ 0.50 m
where d represents the final distance between the predicted trajectory endpoint and the target goal.
Strict success rate across all five models and three goal configurations:
N = 200 samples × 3 goals = 600 evaluations per model
Identify which command regions are associated with failed trajectory predictions.
The DNN-LNA models receive five inputs: GoalX, GoalY, θ, v and ω.
For Experiment 1, the goal position was fixed while 200 combinations of θ, v and ω were uniformly sampled across their operational ranges.
Each sampled configuration was evaluated by all five pretrained DNN-LNA models.
For each prediction, the final distance between the predicted trajectory endpoint and the target goal was computed and used to assign a Success or Failure label.
The resulting outcomes were aggregated into risk maps to identify critical operating regions.
- Initial orientation θ near ±π is the dominant failure trigger
- Angular velocity ω has a secondary influence
- Distinct failure regions emerge in the command space
- Model robustness varies significantly across operating conditions
3D Failure Zone Map
Orientation vs Final Distance
Evaluate model behavior under different target positions.
Three representative goals were tested using both strict and soft success criteria.
| Goal | Position | Relative Difficulty |
|---|---|---|
| A | (0.5, 0.5) | Low–Medium |
| B | (1.0, 0.0) | Medium |
| C | (1.5, -0.5) | Medium–High |
- DNN_LNA_closs2 achieved the highest overall performance
- DNN_LNA_closs1 showed the weakest performance
- Goal position affects performance, although its influence is smaller than orientation and velocity
- Some trajectories classified as failures under the strict metric become near-successes under the soft metric
To improve interpretability, a shallow Decision Tree was trained for each model using:
(GoalX, GoalY, θ, v, ω)
as inputs and:
Success / Failure
as labels.
The extracted trees provide explicit and human-readable rules that describe model behavior and identify critical failure regions.
Additional trees for all evaluated models are available in:
results/figures/
The identified risk regions suggest several potential runtime applications:
Commands located in high-risk regions can be modified before execution.
The navigation system can switch to a classical planner when the predicted risk becomes excessive.
Near-success conditions can be detected early and used to trigger user feedback or additional safety mechanisms.
SafeTraj-Experiments/
│
├── notebooks/
│ └── demo_analysis.ipynb
│
├── src/
│ ├── config.py
│ ├── model_utils.py
│ └── evaluate.py
│
├── results/
│ ├── figures/
│ ├── figures_exp1/
│ ├── figures_exp2/
│ └── tables/
│
├── requirements.txt
├── LICENSE
└── README.md
The original DNN-LNA model weights are not publicly available because they are proprietary assets of the REXASI-PRO project.
Users with access to the original models can adapt the loading functions in src/model_utils.py and execute:
pip install -r requirements.txt
python src/evaluate.pyFor a fully reproducible public example, see:
notebooks/demo_analysis.ipynb
@mastersthesis{bathaei2026safetraj,
author = {Bathaei Pourmand, Pouya},
title = {Analysis of Neural Trajectory Prediction for Collision Avoidance in Smart Wheelchairs},
school = {University of Genoa},
year = {2026},
note = {MSc in Computer Engineering (AI), REXASI-PRO Project}
}MIT License.
See LICENSE for details.
The DNN-LNA model weights remain proprietary assets of the REXASI-PRO project and are not covered by this license.