Robotics · SLAM · Autonomy

Arjun S. Kumar

Technical Lead, Robotics

Botsync · Bangalore, India

See my work Download résumé
01

Map

One scan of a real warehouse: 789,841 LiDAR points.

02

Estimate

Factor-graph SLAM on GTSAM iSAM2 keeps the map and the robot consistent.

03

Deploy

The same stack runs on 1,800+ commercial robots.

Skip intro

Abstract

I build and lead the estimation and autonomy stacks behind mobile robots: 3D LiDAR SLAM, factor-graph optimization and localization, now running on 1,800+ commercial robots worldwide. My work takes a problem from the whiteboard (a cost function, a noise model, a proof that it converges) through to fleet deployment, and I lead the engineering teams and cross-functional work needed to make it hold up in production. Over 8+ years this has spanned warehouse forklifts, commercial cleaning robots, micro aerial vehicles and edge perception.

Keywords: factor graphs, SLAM, state estimation, sensor fusion, GTSAM / iSAM2, LiDAR odometry, ROS 2, modern C++

The problem I work on: factor-graph SLAM X⋆=arg minX∑i‖hi(Xi)−zi‖Σi2

01 — Results

Numbers that shipped.

1,800+
robots running my autonomy stack

Production 3D LiDAR SLAM on GTSAM iSAM2, deployed across a commercial cleaning-robot fleet.

weeks → few minutes
to deploy a new site

Online 3D SLAM replaced offline rosbag recording and hand-corrected pose graphs.

8+ yrs
building production autonomy

Warehouse forklifts, cleaning robots, micro aerial vehicles and edge perception.

355★
GitHub stars on an open-source curriculum

Robotics C++ Notes: modern C++ for robotics engineers.

02 — Experience

Eight years of shipping autonomy.

Jan 2026 – present

Botsync

Bangalore, India

Technical Lead, Robotics Oct 2026 – present

  • Promoted to Technical Lead, Robotics, leading the AMR engineering team and the mapping and localization stack.

Staff Robotics Engineer Jan 2026 – Sep 2026

  • Led the AMR India engineering team, owning the mapping and localization stack end to end and serving as the cross-functional point of contact for DevOps, embedded systems, fleet management and customer support.
  • Replaced manual mapping workflows with a real-time online 3D SLAM pipeline, removing the need to record rosbags and hand-correct pose graphs offline, and cutting site deployment time from weeks to a few minutes.
  • Built a full-featured Map Editor for 2D/3D map layers with multi-map merging, coordinate transformation and manual obstacle-layer editing to speed up deployments.
Aug 2025 – present

Volunteer Instructor, Robotics

Entri.App · Remote · part-time

  • Teaching ROS 2, modern C++, SLAM, state estimation, control theory and robotics mathematics to engineers moving into autonomous robotics.
Feb 2022 – Jul 2025

Senior Engineer, Autonomy

LionsBot International · Singapore

  • Built a production 3D LiDAR SLAM system with GTSAM iSAM2 optimization, eliminating vertical Z-drift in sparse environments for reliable navigation across 1,800+ commercial cleaning robots worldwide.
Apr 2019 – Jan 2022

Senior Engineer, Machine Learning & Robotics

Ignitarium Technologies · Kerala, India

  • Built a complete autonomous navigation stack on the TI Jacinto J7 heterogeneous processor, integrating SLAM, path planning and real-time obstacle avoidance.
  • Optimized custom CNN architectures for real-time 2D/3D object detection on resource-constrained ARM and x86 edge targets.
Earlier roles and education 5
Aug 2018 – Apr 2019

Research Associate, Robotics

Nanyang Technological University · Singapore · contract

  • Designed and built a custom micro aerial vehicle with stereo cameras and a Pixhawk flight controller.
  • Benchmarked visual-inertial odometry pipelines including VINS-Mono and ROVIO on an onboard Nvidia Jetson TX2 for GPS-denied flight.
Jan 2018 – Jul 2018

Robotics Engineer

Addverb Technologies · Noida, India

  • Built an RGB-D bin-picking solution reaching an 80% pick success rate using ROS and PCL, integrated with ABB and Universal Robots manipulators.
Jan 2017 – Dec 2017

Robotics Consultant & Master’s Thesis Student

Ernst & Young (EY) · Kerala, India · contract

  • Delivered industrial proof-of-concept perception solutions using Microsoft HoloLens, RGB-D sensors and CNTK.
  • Thesis research reached 89% accuracy in neural-network surface-crack detection.
Jun 2016 – Oct 2016

Mobile Robotics Intern

Ingeniarius Lda · Coimbra, Portugal

  • Developed and validated swarm exploration algorithms in ROS and Gazebo, extending Darwinian Particle Swarm Optimization to reach 3× faster search-and-rescue area coverage than baseline PSO.
2011 – 2017

Education

Amrita School of Engineering · Kerala, India

  • M.Tech, Robotics & Automation (Jul 2015 – Jul 2017)
  • B.Tech, Computer Science & Engineering (Aug 2011 – May 2015)

Live on the floor

From truck to pallet to outbound.

A truck docks and unloads onto the conveyor. A UR-style arm stacks the cartons on a pallet. When it is full, the forklift reverses in, lifts it and carries it to the outbound zone, while the tugger keeps the aisles moving. Same map, same robots, running live behind this page.

03 — Selected work

Built to run in production.

Open any card for the full write-up.

Botsync, 2026 · production

Real-time online 3D SLAM pipeline

Animated preview (mock): Real-time online 3D SLAM pipeline

Replaced manual mapping workflows that required recording rosbags and hand-correcting pose graphs offline. AMRs now generate 3D maps online in real time, cutting site deployment from weeks to a few minutes.

3D LiDARGTSAMiSAM2C++20ROS 2
SYSTEM ARCHITECTURE: FULL-STACK REAL-TIME 3D LiDAR SLAM PIPELINE INCREMENTAL SMOOTHING & MAPPING VIA GTSAM iSAM2 FACTOR GRAPH OPTIMIZATION 1. SENSOR INGESTION 3D LiDAR Stream 10-20 Hz Raw Packets 6-DoF IMU Sensor 200 Hz High-Rate Acc/Gyr Wheel Odometry Differential Planar Encoders Deskewing Pipeline Continuous IMU Interpolation 2. FRONT-END ODOMETRY Voxel Filter & Normals Adaptive Octree Leaf (PCL) Scan-to-Map Matching Point-to-Plane ICP / NDT Covariance Estimation Σ Keyframe Decider Δd > 0.4m || Δθ > 15° Generates Pose Node xᵢ Local Submap Sliding Window K-Nearest Keyframe Buffer 3. GTSAM iSAM2 FACTOR GRAPH Factor Graph Construction x₀ x₁ x₂ BetweenFactorPose3(xᵢ₋₁, xᵢ) iSAM2 Bayes Tree Smoothing • Incremental QR / Cholesky Factor • Relinearization of affected cliques • Zero Z-Drift Elevation Locking Execution: ~5–12ms per keyframe Loop Closure Integration Non-linear pose constraint insertion: BetweenFactor(xₘ, xₙ, Tₘₙ) Full trajectory back-propagation 4. MAP & FLEET Loop Detection Scan Context / Ring Keys Search ICP Fitness Verify Globally Consistent 3D Point Cloud Multi-Floor Align Zero Z-Drift Map 2D Grid Projection Fleet Deployment 1,800+ AMRs Weeks → Few Minutes Online 3D SLAM LOOP CLOSURE FEEDBACK
Figure 2. System architecture: online 3D LiDAR SLAM & GTSAM factor graph. Four stages deployed on industrial AMRs. (1) Sensor ingestion: 3D point-cloud streaming and 6-DoF IMU pre-integration with high-rate deskewing. (2) Front-end odometry: voxel filtering, point-to-plane ICP/NDT scan-to-map matching, covariance estimation and keyframe selection. (3) Back-end iSAM2: BetweenFactorPose3 odometry constraints, zero-Z-drift elevation locks and Bayes-tree incremental updates. (4) Map & fleet: Scan Context loop detection feeding non-linear loop closures into the graph, producing globally consistent 3D maps and 2D costmaps for 1,800+ AMRs.
Botsync, 2026 · internal tool

Interactive multi-layer 2D/3D map editor

Animated preview (mock): Interactive multi-layer 2D/3D map editor

Editor for 2D and 3D map layers supporting map merging, coordinate alignment and manual map cleaning, built end to end to unblock and speed up site deployments.

  • Multi-map merging and coordinate transformation between maps.
  • Manual obstacle-layer editing on top of the 2D and 3D layers.
Map mergingPoint cloudsPCLTransform pipeline
LionsBot, 2022 – 2025 · global scale

Fleet autonomy stack with zero Z-drift

Animated preview (mock): Fleet autonomy stack with zero Z-drift

Production 3D LiDAR SLAM built on GTSAM iSAM2, eliminating vertical Z-drift in sparse environments. Deployed across 1,800+ commercial cleaning robots worldwide.

LiDAR SLAMGTSAMSensor fusionCommercial AMRs
Ignitarium, 2019 – 2022 · embedded

Edge perception & navigation on TI Jacinto J7

Animated preview (mock): Edge perception and navigation on TI Jacinto J7

Complete autonomous navigation stack (SLAM, path planning, obstacle avoidance) on the TI Jacinto J7, with custom CNN architectures optimized for real-time 2D/3D detection on resource-constrained ARM and x86 edge targets.

TI Jacinto J7ARM / NEONCNN optimizationCUDA / TensorRT

05 — Open source

Giving back.

The open-source robotics community gave me the knowledge and support to build my career. Writing and sharing these notes is my way of giving back to the community that made it possible.

Robotics_CPP_Notes: modern C++ for robotics engineers

★ 343 · 37 forks

A practical reference on C++17/20 for robotics software engineers: CMake build setups, memory management, zero-cost abstractions, concurrency and performance-focused data structures.

C++17/20CMakeRobotics architectureZero-copy abstractionsMemory management

In preparation

  • 3D LiDAR SLAM from scratch. A hands-on guide to factor-graph optimization with GTSAM/iSAM2, loop-closure strategies and multi-floor vertical Z-drift.
  • State estimation & sensor fusion. Mathematical foundations and C++ implementations of error-state Kalman filters, IMU pre-integration and visual-inertial odometry.

06 — Methods & tools

The toolbox.

Estimation & mapping
Factor graphs, iSAM2, LiDAR odometry, scan matching (ICP / NDT), loop closure, IMU pre-integration, VIO, Kalman filtering
Programming & ML
C++, Python, PyTorch, CUDA, TensorRT
Robotics software
ROS / ROS 2, OpenCV, PCL, GTSAM, g2o, Gazebo
Systems & hardware
Git, Docker, CI/CD, Linux (Ubuntu), 2D/3D LiDARs, RGB-D cameras, IMUs, Nvidia Jetson (TX2 / Xavier), TI Jacinto

07 — Contact

Let’s talk.

Always glad to talk about SLAM, estimation, fleet autonomy, robotics leadership or open-source work.