Skip to content
View Telschow's full-sized avatar
🎯
Focusing
🎯
Focusing

Highlights

  • Pro

Block or report Telschow

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Telschow/README.md

Daniel Telschow. Autonomy and applied AI. Systems that know when they are wrong.

Technical product and program lead for autonomous systems. Munich.

In my day job I take automated driving functions from requirements through integration and vehicle validation to release decisions, with engineering teams and external partners. These repositories are where I test the questions behind that work: when does adding a sensor, a camera or an AI component make a system worse, and does the system know it?

Results

14 of 16 scenarios where adding a camera made tracking worse, from mosaic-perception 7 of 8 held-out expectations met, the miss is reported, from nav-trust-c2 17.4 metres actual error when the filter claimed 49 centimetres, one synthetic run, from contested-nav 42 architecture decisions on the record, from contested-nav

All results come from simulation or synthetic data, and each project says so.

Featured work

Operator console showing three drones, one flagged not trusted, with a decision waiting

nav-trust-c2: which drone's position can the operator trust?

Mission control for a small team of simulated drones. Each vehicle is watched for GNSS loss and frozen positions, shown as GREEN, AMBER or RED with a reason, and the console recommends landing or re-tasking. Nothing executes without a recorded operator approval. 3-minute demo.

GNSS loss detected in 1.0 to 1.4 s over 12 runs. The first usability test failed with both participants, so I redesigned the console and published the finding.

PX4 / Gazebo / MAVSDK · FastAPI · React / TypeScript · simulation only

Figure: the filter claims it is within 49 centimetres and is 17.4 metres off, one synthetic run

contested-nav: is the filter right about how wrong it is?

A 21-state error-state Kalman filter and evaluation harness for GNSS-denied navigation. With GNSS denied and camera aiding on, the filter claims centimetres while its error is metres, so vision ships off by default.

When I found a flaw in my own benchmark, I retracted the earlier result, fixed the method and set the bar for the next claim before running it. No candidate met the bar, so I closed the question (ADR-0036).

state estimation · fault detection · uncertainty calibration · Python · synthetic and recorded data

Three tracks of one synthetic target: range sensor only, camera added (worse), bias estimated

mosaic-perception: when does fusing sensors make tracking worse?

A deterministic multi-sensor tracking workbench with paired comparisons. Adding a camera made tracking worse in 14 of 16 scenarios, and an ablation traces most of that to one tracker design choice, not to fusion in general.

multi-target tracking · sensor fusion · Python / Streamlit · synthetic

How I work

  • Decisions are written down as architecture decision records, so a team can disagree with them for the right reasons later.
  • Claims are earned by measurement. Numbers in the docs are generated and checked in CI, and misses stay in the headline.
  • I set the bar before the test, and stop when it is not met.
  • AI-assisted engineering with gates, not trust. Most code here is written by an AI coding agent, and each README says so. I set scope and requirements, make the decisions, review and merge. Tests, held-out protocols and security scans keep it honest.
  • Safety and scope are part of the design: human approval by default, documented non-goals.

Python · TypeScript · state estimation · sensor fusion · PX4 SITL · FastAPI · React · RAG and agent evaluation · GitHub Actions

LinkedIn

Pinned Loading

  1. contested-nav contested-nav Public

    Error-state Kalman filter and calibration harness for GNSS-denied navigation. Shows when a filter is confidently wrong. Synthetic data only.

    Python

  2. mosaic-perception mosaic-perception Public

    Does adding a sensor help tracking? A synthetic-world benchmark where it often doesn't.

    Python

  3. nav-trust-c2 nav-trust-c2 Public

    Human-on-the-loop mission control for simulated drones: GNSS-degradation detection, per-vehicle navigation trust, and recommendations that wait for operator approval. Simulation only.

    Python