> For the complete documentation index, see [llms.txt](https://panav.gitbook.io/robotics-handbook/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://panav.gitbook.io/robotics-handbook/authors-projects/authors-projects.md).

# Author's Projects (overview)

Production robotics projects - Polka (multi-LiDAR fusion), GO-SLAM (custom SLAM stack), BARN Challenge (mapless navigation), Intrinsic AI (learning-based control).

## About this section

Most "robotics portfolios" you find online are toy demos: a TurtleBot following a line, a Gazebo world with three obstacles, a Jupyter notebook with a pretrained YOLO. This section is the opposite. Everything here is **real work from real production robotics** - code that ships on warehouse AMRs, SLAM stacks that get benchmarked against KITTI ground truth, optimization papers headed for peer review, and challenge submissions that move global leaderboards.

I'm Pan, a working robotics engineer based in Bangalore. I graduated IIT Patna with a BTech in Electrical & Electronics Engineering in 2025. The projects below are the ones I keep coming back to: each one taught me something about the gap between "this works on my laptop" and "this works on a robot at 3 AM in a warehouse."

The pages here aren't a CV. They're engineering write-ups - motivation, design choices, what I'd do differently. If you find something useful, copy it. If you find something wrong, [tell me](https://panav.netlify.app).

***

## Projects

<table data-view="cards"><thead><tr><th></th><th data-type="content-ref"></th><th data-hidden data-card-cover data-type="files"></th></tr></thead><tbody><tr><td>Polka - Multi-LiDAR Fusion</td><td><a href="/robotics-handbook/authors-projects/polka.md">Polka - Multi-LiDAR Fusion</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-be3114813958e33b1b4b60b97478dd8efbfefef3%2Fpolka-pipeline-demo.gif?alt=media">polka-pipeline-demo.gif</a></td></tr><tr><td>GO-SLAM - Built from Scratch</td><td><a href="/robotics-handbook/authors-projects/go-slam.md">GO-SLAM - Built from Scratch</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-e112da17eddd596d6c520ae9cffe460d6ac50a3c%2Fgo-slam-demo.gif?alt=media">go-slam-demo.gif</a></td></tr><tr><td>BARN Challenge 2026 - Breadcrumb Explorer</td><td><a href="/robotics-handbook/authors-projects/barn-challenge.md">BARN Challenge 2026 - Breadcrumb Explorer</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-ad49ad0813817fbcf84dea7f87156f742a291f85%2Fbarn-challenge-demo.gif?alt=media">barn-challenge-demo.gif</a></td></tr><tr><td>Intrinsic AI for Industry Challenge</td><td><a href="/robotics-handbook/authors-projects/intrinsic-ai-challenge.md">Intrinsic AI Challenge - Cable Insertion</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-5ed575543bbb40bb4ae997748e5a35170662671b%2Fintrinsic-challenge-hero.jpg?alt=media">intrinsic-challenge-hero.jpg</a></td></tr></tbody></table>

### [Polka](/robotics-handbook/authors-projects/polka.md) - Multi-LiDAR Fusion for ROS 2

A single composable node that merges heterogeneous `PointCloud2` and `LaserScan` streams into unified outputs. Per-source filtering, TF2-aligned fusion, optional CUDA acceleration, and IMU-based deskewing with per-source overrides for articulated platforms. Supports both ROS 2 Humble and Jazzy. Built to replace the relay-filter-merge node soup that plagues most multi-LiDAR stacks.

> **Use case:** Production AMR platforms with 2+ LiDARs, mixed 2D/3D, mast-mounted + base-mounted on a yaw-articulated chassis.

***

### [GO-SLAM](/robotics-handbook/authors-projects/go-slam.md) - SLAM From Scratch (GICP + Pose Graph)

A complete SLAM system I built from scratch in two months: GICP front-end, pose-graph back-end, loop closure, and **custom Levenberg-Marquardt solvers** for both alignment and global optimization - no Ceres, no g2o, no GTSAM. Integrated with ROS 2 Humble using deskewed LiDAR from Polka. Benchmarked on KITTI sequences against ground truth.

> **Not** the academic GO-SLAM paper by Zhang et al. (ICCV 2023). Same name, independent project.

***

### [BARN Challenge 2026](/robotics-handbook/authors-projects/barn-challenge.md) - Mapless Navigation, Solo Submission

IEEE ICRA BARN Challenge 2026 entry. Clearpath Jackal in Gazebo, 270° sensor coverage, dynamic obstacle fields. I built a **Breadcrumb Explorer** architecture instead of the standard SLAM-based baselines - no map, no laser odometry, just an odom-frame memory of "tasty" vs "stale" trajectories that improves over repeated trials. First submission scored **0.3682/0.5** - the highest score by an Indian team since the benchmark began in 2022. Physical finals in Vienna.

***

### [Intrinsic AI for Industry Challenge](/robotics-handbook/authors-projects/intrinsic-ai-challenge.md) - Learning to Plug a Cable

Intrinsic's (Alphabet) AI for Industry Challenge: autonomously plug a fiber-optic cable into a port reachable by a UR5e arm. I climbed the full ladder of approaches - hand-coded FSM in ground-truth mode (280+), classical CV blob detection, learned perception (ResNet18, U-Net), YOLO detection, and finally imitation learning with **ACT** and **SmolVLA** - and hit the real bottleneck: **the data wall**. Scored **286/300** with perfect pose information, but the randomized real-world evaluation (multi-NIC scenes, board-pose randomization) exposed how thin my demonstration coverage was.

> **Lesson:** the constraint was never the model architecture - it was collecting diverse teleop demonstrations. Imitation learning is only as good as its data coverage. Full write-up on [Substack](https://pana1v.substack.com/p/a-journey-through-the-intrinsic-ai).

***

## Reading order

If you're here from my resume or LinkedIn and want the fastest path through this section:

1. Skim [Polka](/robotics-handbook/authors-projects/polka.md) for the **engineering depth** signal (composable nodes, TF2, CUDA, IMU deskewing).
2. Read [GO-SLAM](/robotics-handbook/authors-projects/go-slam.md) for the **fundamentals** signal (custom optimizers, no libraries).
3. Read [BARN](/robotics-handbook/authors-projects/barn-challenge.md) for the **shipping under pressure** signal (solo, leaderboard, finals).

Each page is self-contained. Cross-references are linked where they're load-bearing - Polka feeds deskewed clouds into GO-SLAM, for example.

***

## Find me online

[panav.netlify.app](https://panav.netlify.app) · [github.com/Pana1v](https://github.com/Pana1v) · [linkedin.com/in/panavraaj](https://linkedin.com/in/panavraaj)