I build safe and secure robot autonomy with reinforcement learning, using drones as my testbed. I hold an M.S. in Robotics & Autonomous Systems (AI) from Arizona State University, and I spent two years shipping autonomous mobile robots into live automotive plants. That field work is where my interest in failure-aware autonomy comes from.
Currently
- Research software engineer at Dr. D. Y. Patil College of Physiotherapy, building VR exercise modules and video-based gait analysis for a clinical study on geriatric rehabilitation
- Writing up [M1], learned recovery from partial rotor faults
- [M2] is complete, with arXiv submission pending
- Building a markerless running-gait system for a 120-runner workshop
Note
Robotics researcher in safe & secure robot autonomy, safe RL, and robotics AI security. M.S. Robotics & Autonomous Systems (AI), Arizona State University (GPA 3.73), plus 2+ years shipping autonomous mobile robots in live automotive plants.
Autonomous robots are leaving the lab faster than we can say when they deserve our trust, and I want to help close that gap. I build safe and secure robot autonomy with reinforcement learning, with drones as my current testbed.
|
⚡ Right now
|
💬 Ask me about
|
Every evaluation flight is replayable in 3D on the results site.
|
Drone traffic intelligence YOLO and ByteTrack on 4K drone video produced 35,000+ tracks and 818,000 detections. I georeferenced them from gimbal telemetry and matched them to OpenStreetMap roads at 6.1 m median error for speed, queue and turning analytics. |
DWA local planner for ROS 2 A Dynamic Window Approach planner for TurtleBot3 in Gazebo, with LiDAR collision checking at about 10 ms per cycle. |
|
Markerless running-gait analysis An offline multi-phone system (240 fps capture, RTMPose keypoints, gait-event detection) that reports cadence, contact time, overstride, knee angle and pelvic drop in about five minutes per runner. A comparison study against a D-Wall system comes next. |
Neural inverse kinematics Keras's 74% "accuracy" hid misses larger than half the arm's reach. I traced the cause to inconsistent labels and fixed it: 43× more accurate at the median, 200× on the hardest targets, and a 3.5× faster solver when used as a warm start. |
Graduate projects at Arizona State University
| Project | Summary |
|---|---|
| Lunar Lander ExoCam Earth Analog NASA-affiliated CPS Challenge, 2022 |
A PX4 quadrotor on OptiTrack and an EKF throws a sensor probe outside a perimeter to image its own landing. In phase 2 it delivered the probe onto a moving platform detected by the onboard camera. |
| Mid-flight propeller failure [R1] 2022 |
An RNN fault detector switches between three PPO controllers (4, 3 and 2 working propellers) at 240 Hz in PyBullet, with a PPO vs A2C comparison. This is where my fault-tolerant flight research started. |
| 3D visual navigation with hierarchical RL [R2] 2022 |
An AirSim + PX4 Gym environment and a PPO approach policy trained from RGB (frozen ResNet-18), depth and goal-masked segmentation. |
| Battery-aware scheduling for a LEO CubeSat [R3] 2021 |
CubeSat jobs modeled as priced timed automata with a kinetic battery model, formally verified deadlock-free in UPPAAL CORA. |
| Sky-Hook space tether proposal 2021 |
A technology-readiness analysis, tether scaling, Fusion 360 simulation and business plan for an alternative to rocket launches. |
At Wheel.me in Detroit I deployed robot fleets in live automotive plants and documented the edge cases that broke them. I then built a wheel-slip safety framework that fuses wheel odometry with LiDAR localization, which the General Motors safety team approved. I also built a CAN–ROS–MQTT telemetry framework that cut end-to-end latency by up to 70 seconds.
[M1] D. S. Reddiyar. Learned Mission-Level Recovery from Partial Rotor Faults: Trained in Isaac Lab, Judged with PX4 in the Loop. Manuscript in preparation, 2026. Results
[M2] D. S. Reddiyar. A Post-Processing Adapter for Correcting False Positives and False Negatives in Egocentric Hand Detection. Manuscript complete, arXiv pending, 2026. Code
[J1] P. Shah, A. Gorle, D. Reddiyar, S. Pardeshi, S. Mane. Gait Analysis: Approaches and Applications. IJARIIT 4(3), 204–213, 2018.
Technical reports, Arizona State University
[R1] Control of Mid-Flight Propeller-Deficient Multi-Rotor Using RL. EEE 598 Reinforcement Learning, 2022.
[R2] 3D Visual Navigation with Reinforcement Learning. CSE 591 Perception in Robotics, 2022.
[R3] Battery-Aware Scheduling for a Low Earth Orbit CubeSat. CSE 598 Safe Autonomy for Cyber-Physical Systems, 2021.
[R4] Learning Inverse Kinematics by Imitating a Numerical Solver. Extending MAE 547, 2021. Code
Robotics & simulation
Learning & perception
Research practice
Languages & systems
Awards, courses and languages
Team member, ASU Air Devils (AIAA Design/Build/Fly) · Finalist, TCS CodeVita 2016 and 2017 · Runner-up, e-Zest Megathon 2019 (vision-based rescue with drones)
Understanding Research Methods (University of London) · Deep Learning with TensorFlow 2.0 · Red Hat Certified System Administrator
English, Hindi and Marathi (native) · German (A1, learning)