Robotics Software · Reinforcement Learning · AI/ML B.Tech Mathematics & Computing, Central University of Karnataka (2023–2027)

I work at the intersection of robotics, reinforcement learning, and intelligent control, with hands-on experience in bipedal locomotion, inverse kinematics, motion planning, robot simulation, and learning-based control.

I have worked with MuJoCo and PyBullet, developing controllers and RL systems for both simulated and physical robotic systems.

Currently exploring: Reinforcement Learning · Robot Learning · Optimal Control · MPC · Intelligent Robotics

May 2026 – July 2026

• Implemented PD and model-based controllers in MuJoCo for CartPole, Furuta Pendulum, and Reaction Wheel Pendulum.
• Deployed PD balancing on physical hardware, maintaining upright stability for 10+ seconds under real-world noise and actuator delays.
• Trained a PPO agent with a custom reward to learn bipedal sit-to-stand from IK-generated reference trajectories.
• Developed IK pipelines for deep squat-to-stand, sit-to-stand, and fallen-to-stand transitions with smooth joint trajectories.
• Engineered a PD torque controller for dynamic deep squat-to-stand motion.

May 2025 – December 2025

• Developed 3+ RL-based biped locomotion systems using custom Gymnasium and PyBullet environments.
• Benchmarked SAC, TD3, and DDPG on a 6–8 DOF underactuated biped over 10M+ training steps.
• Achieved 94% navigation success and 2% fall rate under stochastic disturbances.
• Integrated A* global planning with SAC-based locomotion for hierarchical obstacle avoidance, achieving 87.6% goal success in cluttered environments.
• Designed progressive waypoint-based reward shaping for stable, goal-directed locomotion.

October 2024 – December 2024

• Completed 3+ end-to-end data analytics projects involving EDA, feature engineering, and predictive modeling.
• Built predictive pipelines using Python, Pandas, and Scikit-learn, achieving 75–85% model accuracy.

Published — IEEE ICC 2025

Automated pipeline using Gemini 2.5 Flash to iteratively generate and refine reward functions for a SAC-trained biped across flat, uneven, slope, and stair terrains.

• 151+ unit course completion vs. −0.76 for the vanilla PPO baseline.
• Gait symmetry index of 0.023 and torso tilt of 0.086 rad.
• Evaluated 21 gait-specific metrics.
• Integrated human-guided constraints for naturalistic locomotion.
• Implemented automated training resume and rollback on generated-code failures.

Gemini 2.5 Flash SAC PyBullet Gymnasium NumPy Matplotlib

6–8 DOF underactuated biped trained with SAC, TD3, and DDPG over 10M+ steps, combined with A* hierarchical planning and waypoint-based reward shaping.

94% navigation success · 2% fall rate

PyBullet SAC TD3 DDPG A* Gymnasium

Developed IK-generated reference trajectories, PPO-based learning, and PD torque control for dynamic biped stand-up transitions.

MuJoCo PPO Inverse Kinematics PD Control

Real-time contactless attendance system with 128-D face encoding, liveness detection, and a secure web dashboard with live camera streaming and CSV export.

Python Flask OpenCV dlib ONNX SQLite

Fully connected neural network implemented using NumPy with manual forward/backward propagation, softmax, cross-entropy loss, and mini-batch gradient descent.

NumPy Matplotlib

Real-time gesture-controlled drawing system supporting drawing, erasing, brush control, and color selection.

OpenCV MediaPipe NumPy

Languages Python · C++ · JavaScript · SQL · R

ML / Deep Learning PyTorch · TensorFlow · Keras · Scikit-learn · NumPy · Pandas

Robotics / RL / Simulation Gymnasium · PyBullet · MuJoCo · MjLab · SAC · PPO · TD3 · DDPG · Inverse Kinematics · Motion Planning · PD Control

Computer Vision OpenCV · MediaPipe · dlib · ONNX

Tools Git · TensorBoard · Flask · SQLite · Power BI

Open to internship opportunities in Software Robotics, ML Engineering, and Data Science, as well as research collaborations.

📧 abhaydwivedi10122005@gmail.com