Robotics · Reinforcement Learning · Graph Neural Networks · Sim-to-Real Control

B.Tech Computer Science and Engineering (AI & Data Science) · MIT-WPU, Pune · Expected 2027

Research Intern, CAIR, IIT Mandi · RL-based quadruped locomotion.

Building across AI/ML, robotics, deep learning, software systems, and embedded engineering. Open to AI/ML, robotics, software, and embedded systems internships from January 2027.

LinkedIn · Email · GitHub Repositories

Centre for Artificial Intelligence and Robotics (CAIR), IIT Mandi · Research Intern, July–December 2026

Developing a hierarchical reinforcement-learning locomotion and mapless obstacle-avoidance pipeline for the 12-DOF Hiwonder ROSpug quadruped in NVIDIA Isaac Lab, focused on actuator latency modeling, state estimation, and real-time Jetson Nano deployment.

• Hierarchical Locomotion & Obstacle Avoidance: Built a two-tier RL controller pairing a 20 Hz omnidirectional Hopf-CPG + PPO primitive (analytical kinematic steering/crabbing prior; 100% survival across 8 velocity regimes) with a 4 Hz 2D-LiDAR PPO avoidance policy (2.5 m range parity, OpenUSD warehouse clutter, and 3-sided U-trap escape shaping).
• System Identification & Latency Modeling: Executed a 480-trial step-response sweep across all 12 serial-bus servos, identifying a 59 ms base latency alongside duration-dependent dynamics; integrated Actuator-Net models and circular FIFO command-delay buffers in Isaac Lab.
• Software State Estimation: Engineered a calibrated 2nd-order LTI joint observer (ωₙ = 22.5 rad/s, ζ = 0.89) to bypass 241 ms sequential bus polling, reducing held-out joint prediction error by 50.6%–62.1% (p = 0.031) over nominal 1st-order filters.
• Hardware Validation & Edge Deployment: Deployed dual ONNX Runtime inference (0.34–0.81 ms) and 36-sector LiDAR filtering with a non-blocking telemetry pipeline on Jetson Nano, holding 50.0 ms loop periodicity across 36 runway trials (0 falls) and 148 s closed-loop obstacle-avoidance runs.

Research manuscript in preparation.

Investigating zero-shot locomotion policy transfer across robot morphologies using a 31,582-parameter heterogeneous Graph Neural Network (GATv2Conv) trained with Proximal Policy Optimization (PPO). The system combines graph-based control with YOLOv8 perception and an on-device LLM planning layer, with ROS 2 Jazzy and Gazebo integration.

• Zero-shot quadruped-to-hexapod transfer: Transferred a policy from a 12-DOF quadruped to an 18-DOF hexapod without retraining, obtaining a reward of 106 ± 25 and approximately 47 survival steps. The fixed-input-dimension MLP baseline could not run on the unseen morphology.
• Fine-tuning: 500K additional steps increased reward from approximately 110 to 416 ± 114 and survival from 47 to 193 steps.
• Model size: 31,582 parameters, approximately 85% fewer than the 210,457-parameter MLP baseline. The MLP achieves higher in-distribution reward, while the GNN supports variable graph structures.
• Terrain evaluation: 95% success at a 5° slope and 0% at a 10° slope in the reported zero-shot terrain tests.
• Control and deployment: Reported 200 Hz GNN inference with yaw-rate PI correction to address circular drift associated with HAA joint offsets.
• LLM planning: Qwen 2.5 7B through Ollama maps natural-language instructions to skills and GNN-generated joint commands, running locally without cloud inference.

Limitation: Zero-shot transfer is not yet robust across all conditions. The hexapod's initial survival is substantially below the in-distribution result, and the 10° terrain result shows that the current policy has clear limits.

Repository, implementation details, results, and reproduction instructions · Zenodo record

A local-first Linux desktop AI daemon built around persistent services, UNIX domain socket IPC, X11/OCR context tracking, and streaming LLM inference through Ollama, with optional online fallback.

• Implemented an intent router that dispatches requests to typed executors, including summarization and error explanation.
• Built a fault-tolerant streaming pipeline that separates the CLI client lifecycle from the persistent daemon.
• Added internet-aware backend switching without requiring a daemon restart.

Medical-image polyp segmentation model combining attention gates, squeeze-and-excitation (SE) blocks, and BCE-Dice loss. Evaluated on Kvasir-SEG, with a reported Dice score of 0.9436.

PPO-based reinforcement-learning pipeline in MuJoCo for grasping and lifting a cube with the LEAP Hand. The repository reports an 85% success rate across 100 randomized deterministic evaluation episodes. The policy meets the defined binary success criterion, but sustained grasp quality remains a limitation documented in the repository.

Embedded C firmware for Team Vegapod, including a custom Bluetooth bootloader, FreeRTOS-based BMS/VCU braking, three-phase inverter PWM, and CAN communication. Represented MIT-WPU at European Hyperloop Week 2025 in the Netherlands.

• ResNet-50 from Scratch: Implemented ResNet-50 in PyTorch and reported 86.63% top-1 accuracy on Imagenette, with live ESP32-CAM inference.
• Brain Tumor MRI Classification: Adapted a U-Net encoder for binary MRI classification, reporting 98% accuracy on a dataset of 3,000+ images.

Investigated a reported visualizer-configuration crash and traced it to a broken workspace installation rather than an underlying logic bug. The investigation surfaced stale visualizer-alias assertions and a test for a CLI flag removed in version 3.0. The resulting test fixes were merged in September 2026 and backported to release/3.0.0.

Fixed a schema-validation edge case involving zero-sized validation splits in ML training pipelines. The pull request passed 17 CI checks and was merged in January 2026.

Robotics and RL

Machine Learning and Deployment

Systems and Embedded

• LinkedIn: linkedin.com/in/chaitanyaparate
• Email: paratechaitanya5@gmail.com
• GitHub: github.com/ChaitanyaParate