Experience

Robotics, simulation, and machine learning roles — industry and research.

Detailed timeline of my professional and research positions. For portfolio-style write-ups, see Work, Research, and Personal projects.

Richtech Robotics USA

Robotics Engineer · Dec 2025 – Present · Las Vegas, NV

  • Ported Richtech’s robot fleet into Isaac Sim — dual-arm manipulators with grippers and mobile robots without hands — with consistent URDF/CAD import, collisions, and reusable USD workflows. Project →
  • Built a bimanual Rubik’s cube task: right arm picks, mid-air handoff to left arm, place on table; ~2k episodes via Isaac Lab multi-environment / vectorized generation; π0.5 fine-tuning with strong sim-to-real on hardware. Project →
  • Created a packing-scene synthetic dataset (similar bimanual platform) with a deformable bag in sim — high-difficulty soft / contact-rich manipulation; includes a polished demo video on the project page. Project →
  • Trained a PPO + diff-IK reach policy with obstacle avoidance on the same bimanual platform — moved position error from a stuck ~18-mm plateau down to ~5 mm by adding a non-saturating threshold penalty above the saturated-tanh equilibrium; characterized the position-vs-orientation Pareto and the PPO stability ceiling observed on this task. Intended to replace planner-based motion in the packing pipeline. Project →
  • Implemented VR teleoperation for dual arms and a hand gripper to grip diverse objects and support data collection on real systems (ROS where needed). Project →
  • Ongoing: PhysX tuning (contact, friction, deformables), cuRobo-style planning integration where applicable, Blender mesh prep for sim assets, and large-scale imitation / policy-training dataset tooling beyond the projects above.

DL-RL

Robotics Engineer Intern · Apr 2025 – Dec 2025 · Brooklyn, NY

  • Built large-scale synthetic datasets in Isaac Sim for the SO-100 and Trossen arms using IK-driven rollouts, domain randomization, and multi-sensor pipelines; published datasets on Hugging Face (1k+ monthly downloads).
  • Developed end-to-end Isaac Sim / Omniverse workflows: scene and physics setup, USD stage management, CAD/URDF integration, camera rigs, and Replicator-based auto-annotation for fully labeled synthetic data.
  • Designed and ran robotics simulation tasks — manipulation, grasping, and pick-and-place — to validate policies before real hardware deployment.
  • Fine-tuned Gr00t and π0 models in PyTorch on synthetic and real data; closed-loop testing on the physical SO-100 arm with 90–100% task success on benchmark routines.

Project write-up: Benchmarking Gr00t & sim-to-real on the SO100 arm.

Stony Brook University

Graduate Research Assistant · Sep 2020 – Dec 2025 · Stony Brook, NY

  • Designed and trained multi-modal LLM-style models with Mixture-of-Experts (MoE) for path synthesis, built from scratch with a hybrid ViT + decoder architecture; CLIP-style contrastive learning and Gaussian soft objectives, with ~15% accuracy gains over strong baselines at ~200 ms inference latency.
  • Improved training efficiency with Classifier-Free Guidance and LoRA (~30% compute reduction); gradient tracking, adaptive clipping, and CLIP-aligned objectives for cross-modal consistency.
  • Developed β-VAE and graph-based VAE models for structured latent representations supporting multi-modal generation and downstream reasoning.
  • Co-authored ASME JMD work on a 3M-sample planar linkage dataset for ML-driven path synthesis and ASME JMR work on conditional β-VAE path generative models for mechanism design.
  • Ran systematic studies comparing curve representations (Fourier descriptors, wavelets, point coordinates, images) inside unified generative frameworks for four-bar coupler synthesis.
  • Led end-to-end ML work: curation and preprocessing of 12M+ multi-modal samples; distributed training of sub-1B transformer models on SLURM-managed HPC; scaling experiments on cloud GPUs; 200+ experiments tracked with PyTorch Lightning, Hugging Face, and Weights & Biases.

Zortag

Computer Vision Engineer · Aug 2024 – Aug 2025 · St. James, NY

  • Fine-tuned YOLOv11, pushing detection accuracy toward ~100% and cutting latency by replacing a two-step pipeline with a single optimized model.
  • Automated myCobot 280 PI capture workflows (~60% less manual work, ~3× labeling throughput) with automated AWS S3 uploads for dataset growth.
  • Shipped a real-time iPhone QR detector (SwiftUI + CoreML), containerized with Docker for scalable deployment.
  • Shipped live screen vs. non-screen classification in the iOS product: CLIP ViT-L/14@336 + MLP (after YOLO proved unreliable on real displays—glare, moiré, weak box boundaries), with WebSocket camera streaming, 5-crop TTA, EMA, and 3-frame hysteresis for stable ~sub-second inference.
  • Developed training utilities with albumentations, batched GPU feature extraction, stratified splits, and disk-level caching (scikit-learn, PyTorch) for fast iteration across classifier architectures.

Related portfolio page: Dataset automation with myCobot & iPhone.