ATHARVA PATWE

• Building and leading a team of Reinforcement Learning engineers, driving research and development of locomotion and whole-body loco-manipulation policies for the 4NE1 humanoid. • Developing a comprehensive skill library for the 4NE1 humanoid, including locomotion capabilities such as walking, running, and stair climbing, along with whole-body loco-manipulation tasks using Reinforcement Learning. • Leading full-stack sim-to-real development, including simulation asset modeling, actuator-level and system-level system identification with simulator integration, environment design, and end-to-end RL policy training. • Developing and maintaining scalable training and learning frameworks supporting diverse RL algorithms, multi-GPU distributed training, and rapid policy iteration cycles. • Leading hardware-software stack integration with Qualcomm chipsets, enabling efficient on-device inference for real-time deployment on the 4NE1 humanoid. • Architected a modular, production-grade C++ inference pipeline using FastDDS, ONNX Runtime, and TensorRT, featuring a configuration-based API for no-code policy deployment, safe operation, and seamless sim-to-real and inter-policy transitions.

• Developed a three-phase RL framework for the 32-DOF 4NE-1 humanoid using PPO with a privileged Oracle policy in Isaac Sim, achieving robust locomotion via domain randomization for sim-to-real transfer. • Implemented an attention-guided world model with a multimodal encoder-decoder, processing proprioceptive and exteroceptive heightmaps via CNNs and multi-head attention LSTMs for state prediction and distillation. • Led full-stack sim-to-real development pipeline: from URDF/USD modeling, actuator/friction tuning, environment design, and RL policy training to inference pipeline. • Integrated state estimation into the FastDDS pipeline, enabling communication with a real-time EtherCAT controller via ONNX-based inference. Achieved accurate velocity tracking and natural gaits on flat/rough terrain and during object transportation. • Developed a comprehensive multi-agent simulation framework that supports various robot embodiments and physics simulators (Sim2Sim), featuring various RL agents, sim-to-real pipelines, and distributed training infrastructure for locomotion, navigation, and manipulation tasks.

◦ Selected as one of 25 scholars from the pool of 1000 applications to carry out fully funded Master's Studies in Europe

◦Engineered a robust RRT-based path planner plugin and conducted extensive real-world effectiveness analyses

-Developed a Computer Vision-driven system for inspecting Engine components employing Deep CNNs -Reduced expenses up to 85% while retaining an accuracy of 96.5%, and the project is now being implemented in all CVBU plants across INDIA -Implemented an Explanation framework for Non-Linear Classifier Decisions using Layer-wise Relevance Propagation -Conducted data analysis and initiated a predictive maintenance model, leading to a significant decrease in downtime and a 20% improvement in overall equipment efficiency (OEE). -Built a data augmentation strategy for image datasets, improving model robustness and generalization capabilities

◦Designed, analyzed, simulated, and manufactured a chassis, applying concepts of vehicle dynamics and handling ◦Developed Data Acquisition System for the FSAE vehicle Based on CAN Bus Communication Interface ◦Engineered an end-to-end Deep Reinforcement Learning(RL) control architecture proof of concept for an autonomous Formula SAE vehicle ◦Led a group of 12 members in the Cost and Manufacturing Static event at Formula Bharat 2022 and achieved all India 2nd position in this event Constructed a modular redundant subsystem integrating robust EKF-based State Estimation, LiDAR & Visual Inertial Camera setup-based perception, and particle filter-based SLAM Developed a Model Predictive Control strategy for following race line, which includes dynamic modeling, constraints, and the real-time optimization solver

· Collaborated with ONGC’s robotics team to develop and implement innovative solutions for onshore production and processing facilities · Implemented reinforcement learning (RL) algorithms, such as DDPG, SAC, to train legged robots for autonomous navigation and obstacle avoidance in dynamic environments encountered during inspection missions · Integrated perception sensors, such as LiDAR and RGB-D cameras, into the robotic platform, to enable real-time environment mapping and localization, enhancing situational awareness for efficient path planning.