Buttcloud

A private research vault for my PhD research on fallback path planning under camera degradation.

The thesis asks whether end-to-end driving models can still extract task-relevant features when camera input is severely corrupted, then use runtime safety supervision to make fallback decisions that can be audited later.

PhD thesis End-to-end AI for Automated Driving: Fallback Path Planner under Camera Degradation with Runtime Safety Supervision for Auditable Fallback Decisions

PhD candidate Ing. Lemnariu Flaviu Rares Technical University of Cluj-Napoca

PhD coordinator Prof. Dr. Ing. MUNTEAN Ionut Technical University of Cluj-Napoca

Collaborator Assoc. Prof. dr. ing. MORRIS Brendan University of Nevada

View repository

Current status MVP preparation phase

The project has completed the initial definition work and is moving into dataset generation, fallback trigger logic, and the first monitoring interface.

NowDataset v0degradation recipe, Apr-Jun 2026
NextFallback triggernominal to fallback switching, May-Jun 2026
MVPLive/history UImonitoring interface, Jul-Aug 2026
SelectCandidate memoarchitecture selection, Aug-Sep 2026

How the site is used

Sign inUse the Access Workspace button to open the login dialog.
InspectViewers can browse loaded scenarios, preview camera evidence and analyze research results.
ContributeYou can ask for editor rights via e-mail to contribute@rlemnariu.work. Then you can upload datasets or generated sequences.

Some technical aspects

Object storageRaw assets are stored in an S3-compatible bucket.

Metadata databaseFile records, owners, roles, and audit context are stored separately.

Service architectureAPI Gateway, Storage, Metadata, and Compute services separate authentication, object storage, database records, and GPU orchestration.

CI/CD pipelineAutomated checks reduce regressions before deployment.

Cost controlsOperational limits keep storage and compute usage predictable.

Research question

Can neural networks extract sufficient task-relevant features from severely corrupted camera input, then support a safe fallback decision instead of simply handing control away?

Gap

Current approaches often depend on LiDAR, perfect real-time HD maps, redundancy, or conservative clock-out behavior; usable feature extraction from degraded camera images remains missing.

Methodology

Use Alpamayo as the nominal driving baseline and fallback trigger, Cosmos for synthetic edge cases, plus Bosch and open datasets for grounding and validation.

This workspace

Supports the MVP interface for live and history monitoring: storing visual sequences, previewing evidence, managing ownership, and auditing role-based access.

Build 0724899 · Jun 6, 2026 02:41 EEST

Select a sequence

Directory Log

Buttcloud

Select a sequence

Directory Log

Raw SQL Executor

Identity Management

Compute Analytics