Mila · MIST Lab · 2025–2026 · First-author preprint
AURA*: planning under perception uncertainty
[One-sentence summary: what it is, who it is for, and the main result.]
Context
[Where and why the project started: employer, course or competition, and the people involved.]
[The constraints: deadline, data available, tools imposed, confidentiality.]
Problem
A planner that treats every uncertain cell as forbidden stops moving. One that ignores uncertainty takes risks it cannot see.
Approach
A path planner that turns calibrated perception uncertainty, from 2D imagery and 3D LiDAR, into a continuous risk field and searches it with an adaptive A*.
- [Design decision 1, and why you chose it over the alternative]
- [Design decision 2, and why]
- [Design decision 3, and why]
Results
100% mission success where the geometric baseline falls to 0.4%. Maximum epistemic risk down 94.9% for about 9% extra path length.
[How it was evaluated: data, baseline, protocol, and who validated it.]
Limits and next steps
- [What does not work yet, or the known limits of the result]
- [What you would do differently]
- [The next step, if the project continues]
Stack
- Python
- PyTorch
- SegFormer
- PointNet++
- conformal prediction
- [Add a tool]