Skip to content

Projects

A curated set of End-to-end robotics projects across Perception & SLAM, Industrial systems, Controls/Mechanisms, and Learning-based control.
Each card lists what I owned, the stack, and one outcome—with deep dives where useful.

  • Perception & SLAM --- 6DoF Pose Estimation, LiDAR SLAM, ORB-SLAM3 + YOLO
    View

  • Industrial Systems --- AGV + Pick-to-Light, Stock Mgmt (SAP)
    View

  • Controls & Mech --- UR5 IK, Base-actuated RCM
    View

  • Learning-based Control --- Deep RL Navigation (DDPG/TD3)
    View


Perception & SLAM

Perception Model for 6DoF Pose Estimation

Summary. Built a baseline multi-view pipeline for 6DoF pose of industrial parts: YOLOv11 detects, ResNet50/SimplePoseNet regresses rotation per view, and epipolar matching + triangulation recovers translation, evaluated on IPD (Industrial Plenoptic Dataset).
My role. End-to-end pipeline, training, epipolar matching/triangulation, analysis.

  • Design: Decouple R (learned per-view) and t (multi-view geometry); Hungarian matching on symmetric epipolar distance before triangulation.
  • Software: YOLOv11 (Ultralytics), PyTorch ResNet50 PoseNet, calibrated 3-cam rig.
  • Metrics/notes: Report ADD-S and rotation error; detector mAP≈0.97 (synth), 0.84/0.81 P/R on real; 3-view runtime ≈ <50 ms on RTX 4070.
  • Outcome: Clean geometric t with consistent multi-view matches; identified next steps (ICP refinement, direct t regression, correspondence-based upgrades).

Read more →


GNSS-aided 3D LiDAR SLAM (Lego-LOAM + GPS)

Summary. Extended Lego-LOAM with an absolute GPS unary factor in a factor graph to control drift—especially in the z-axis—on long, real-world routes. Evaluated on UTBM and Mulran datasets with clear drift reduction versus baseline. My role. GPS factor integration & tuning (noise/gating), factor-graph plumbing in GTSAM, dataset evaluation/plots, and write-up.

  • Architecture / design: Classic LOAM front-end (odometry + mapping threads) + loop-closure; iSAM2 back-end with GPSFactor fused as an absolute prior (LLA → local Cartesian via GeographicLib); outlier handling when GPS “jumps”.
  • Software: Built on Lego-LOAM, influenced by LIO-SAM and hdl_graph_slam; back-end in GTSAM (no custom Jacobians needed vs Ceres).
  • Methods: Converted GPS to ENU/local frame; adaptive noise and covariance checks; delayed trust on re-initialization; factors added incrementally before iSAM2 updates.
  • Outcome: Lower global drift on long sequences; notable z-drift reduction; robust behavior in multi-environment runs.

Read more →


Learning-Based Feature Selection for ORB-SLAM3

Summary. Integrated YOLOv5 with ORB-SLAM3 to drop ORB features that fall inside dynamic-object detections, aiming to stabilize data association in real scenes. Evaluated on EuRoC, KITTI 00, and a custom NUance driving dataset.
My role. Integrated the YOLO→feature filtering step, ran experiments, and analyzed trajectories/metrics.

  • Architecture / design: Keep the standard ORB-SLAM3 pipeline; after ORB keypoints are extracted, run YOLOv5 and remove features inside dynamic-class boxes (person, car, truck, etc.). Tracking, mapping, and loop-closure proceed unchanged.
  • Software: ORB-SLAM3 (C++), YOLOv5 for detections; scripts for plotting trajectory overlays and L1 error against GT/GPS.
  • Methods: Compare baseline vs. filtered variant across datasets; qualitative overlays + L1 absolute error.
  • Outcome: EuRoC: near-perfect overlays. KITTI 00: parity with vanilla ORB-SLAM3 (~0.5% drift). NUance (~2 FPS): cleaner straight-line tracking; sharp turns still fail due to too few frames. Also compared with DROID-SLAM as a learning baseline (strong accuracy, high compute).

Read more →


Learning-based Control

Autonomous Robot Navigation with Deep RL

Summary. Trained a differential-drive robot in a custom 2D “EscapeRoom” simulator to reach a goal without collisions using DDPG and TD3. Continuous actions are left/right wheel speeds; a shaped reward balances progress, heading, time, collisions, and success bonus.
My role. Co-author: environment + reward shaping, PyTorch implementation, training/eval, analysis.

  • Design:
    State: [x, y, heading, linear & angular speed, distance to goal, angle to goal]
    Action: [left wheel speed, right wheel speed] ∈ [-1, 1]
  • Algorithms: DDPG (actor–critic with replay + targets) and TD3 (double critics, delayed policy, target smoothing).
  • Outcome: DDPG reduced collisions but showed critic instability; TD3 stabilized learning and improved episode reward trends, though occasional regressions remained.

Read more →


Controls & Mechanisms

UR5 — Kinematic Control for a Square Trajectory

Summary. Programmed a UR5 to trace a square path in 3D while keeping a fixed tool orientation. Built the forward kinematics from a DH model and used a Newton/iterative IK (Jacobian-based, body frame) to solve \( \theta \) for each waypoint; added a 3D animation to verify the motion.
My role. End-to-end: DH modeling, FK, IK solver, trajectory planner, and animation in MATLAB.

  • Robot / model: UR5, 6R manipulator; standard DH table and link lengths per spec.
  • FK: chained transforms to obtain the end-effector pose \(T^0_6\).
  • Trajectory: corner waypoints define a task-space square; interpolate line segments.
  • IK: Newton’s method with the Jacobian; iterates from the previous pose for smooth convergence.
  • Validation: animated stick-figure UR5 following the path; orientation held constant.

Base-Actuated 3-Rhombus RCM Mechanism

Summary. Designed a planar 2-DOF Remote Center of Motion (RCM) mechanism with all actuators on the base to avoid proximal-actuation issues (RCM offset, added inertia, blocked field of view). Work covers concept → kinematics → singularities → SolidWorks simulation; intended for MIS tasks (e.g., lumbar puncture).
My role. Solo project: literature review, mechanism synthesis, analysis, CAD and motion verification.

  • Architecture / design: Three-rhombus linkage; two active links at the base drive the end-effector. Decoupled motions: same-direction drive → rotation about RCM; opposite-direction drive → translation.
  • Analysis: Planar DOF check (Grübler) confirms 2 DOF.
    Forward kinematics: θ = ½(θ₁ + θ₂) , d = L₅ − L₃·cos(α) − L₁·cos((θ₁ − θ₂)/2).
    Singularity when α = (θ₁ − θ₂)/2.
  • Implementation: SolidWorks model with base motors + linear guides; needle as end-effector; verified home/extended and ±rotation configs.
  • Outcome: Validated RCM behavior and workspace sector; promising for MIS; next step: spatial extension.

All-Terrain Vehicle (ATV) — SAE BAJA (Team Ezhal)

Summary. Competition-grade ATV from concept to race; limited budget, high scrutiny.
My role. CAE lead (Ansys), transmission support, and fabrication contributor.
- Design compliance: Full CAD to SAE BAJA rulebook; crash/impact cases simulated and presented in first-round review.
- Fabrication: Machining, welding, assembly, testing with rapid design changes in electrical and drivetrain.
- Innovation: Disengageable driveshaft for tighter maneuverability and controlled drifting.
- Reliability fixes: Mud-blocking inlet cover post-tests to protect intake.
- Result: State Runner-ups .

Fabrication and assembly ATV frame CAD Ansys crash/impact setup Competition test runs

Industrial Systems

Autonomously Guided Vehicle (AGV) — Accenture (Bengaluru)

Summary. Three-wheeled AGV (two powered steppers + caster) integrated into a warehouse flow; edge control with cloud coordination.
My role. End-to-end prototyping (mechanical, electronics, firmware) and system integration.
- Architecture: Raspberry Pi as the central controller; two Arduinos running PID loops for each stepper; custom motor driver for current control.
- Software: MongoDB + Node.js + React; a local server on the Pi syncs with a central server and inventory DB.
- Navigation: Pre-fed location matrix and odometry; tuned for smooth starts/stops and accurate docking.
- Outcome: Predictable task execution on the shop floor with live status and hand-offs.

AGV prototype – chassis and electronics bay AGV – dual stepper drive with custom driver

iWarehouse — AGV + Pick-to-Light Integration (Accenture)

Summary. Merged Pick-to-Light with the AGV for autonomous putaway/replenishment.
My role. Hardware–software integration and workflow design.
- Pick-to-Light: Weight-based bin sensing triggers light cues; count displays show current inventory and pick/put confirmations.
- Flow: Orders → central server → Pi (edge) → AGV motion + light cues; confirmations sync back to inventory.
- Tech: Same AGV control stack; expanded React UI for operators.
- Outcome: Faster material movement and fewer pick errors; compelling demo for stakeholders/customers.

iWarehouse – bin weights and count display iWarehouse – pick-to-light lanes with AGV

Intelligent Stock Management System — Accenture (Chennai)

Summary. IoT stock monitoring with load-cell bins that auto-initiate refill orders; demo integrated with SAP.
My role. Hardware design (sensor + power), firmware, and backend integration.
- Sensing: ESP8266 nodes read load cells; thresholding/smoothing to reject noise and detect real consumption.
- Edge → server: Local Pi broker relays updates → central server → SAP for order creation in the demo environment.
- Electronics: Linear regulation, grounding/decoupling for stable ADC readings.
- Recognition: Best Innovator Award (Accenture, Oct 2019).
- Outcome: Reliable refill triggers and bin-level visibility; reduced manual checks and out-of-stock risk.

Load-cell bin node with ESP8266 Stock dashboard and refill trigger flow

Hobby/Fun Projects

Robotics & Automation

KUKA Robotic Arm Drawing

Summary. Parametric motion pipelines for KUKA using Grasshopper integrated with KRC.
My role. Toolchain integration and workflow setup.
- Built Grasshopper definitions to generate toolpaths and poses; exported to KRC programs.
- Reduced iteration time from CAD geometry to robot motion with reusable templates.

Grasshopper-driven KUKA path planning

Mitsubishi Robotic Arm Milling

Summary. Added a high-torque spindle to a Mitsubishi arm for light milling operations.
My role. Integration and controls.
- Robot control: Mitsubishi Melfa for arm actuation; GX Works for PLC I/O/safety.
- Mechatronics: Fixture design, stiffness checks, conservative feeds/speeds for proof-of-concept.
- Outcome: Successful pilot milling; defined requirements for rigidity upgrades.

Mitsubishi arm with spindle for milling

IoT

Smart Room

Summary. Custom PCB + firmware to retrofit a standard room with IoT controls.
- Electronics: KiCad board around an ESP module; relay drivers, power conditioning, isolation.
- Firmware: Wi-Fi provisioning, topic-based control, simple REST/MQTT endpoints.
- Outcome: Reliable switching/telemetry with safe mains isolation.

Smart Room PCB and control panel

Smart Locker

Summary. Office-space locker system for guided deliveries and pickups.
- Compute: Raspberry Pi with camera and simple UI.
- Mechanics: Solenoid/servo latch control; status LEDs; QR-based pickup flow.
- Outcome: Reduced hand-offs and better audit trail for small-parcel movement.

Smart Locker – external view Smart Locker – internal mechanism

Miscellaneous projects

Smart India Hackathon(2019)

Summary. Knowledge Management portal for DPSUs to share practices and avoid repeat mistakes.
My role. Front-end implementation and UX refinement.
- Stack: React, Node/Express, PostgreSQL.
- Features: Discussion threads, resource repositories, searchable case histories.
- Outcome: 'Won the 2019 edition'; demonstrated tangible value in cross-org knowledge reuse.