1) Stood up in-house foundation model for multi-modal perception with emphasis on 3D reasoning2) Achieved generalization to various downstream tasks such as depth-from-stereo and robotic manipulation3) Demonstrated neural scaling laws (to 100B tokens) for self-supervised pretraining 4) Set up and maintained cloud GPU cluster with high-speed interconnect

Tech lead for:1) Learning-based methods for 2D & 3D synthetic data generation2) Multi-modal synthetic data pipeline for tasks such as: depth-estimation, segmentation, robotic manipulation3) Scalable, high-throughput cpp renderer optimized for large-scale procedural synthetic data generation

• Robot perception and manipulation. My work involves building and training deep neural networks, researching new algorithms, and building the technical infrastructure to support working with large datasets.

• Robot perception and manipulation. My work involves building and training deep neural networks, researching new algorithms, and building the technical infrastructure to support working with large datasets.

• Robot perception and manipulation. My work involves building and training deep neural networks, researching new algorithms, and building the technical infrastructure to support working with large datasets.

• Worked on the SOTA Dexterity Network (Dex-Net) project (http://berkeleyautomation.github.io/dex-net/) for robust robot grasping using extensive domain randomization enabling sim2real transfer and robustness to imprecision in sensing and control.• Developed the GQ-CNN (https://github.com/BerkeleyAutomation/gqcnn) framework for efficiently training and deploying deep neural networks to plan robust robot grasps.• Worked on grasp-optimized motion planning using learning-based methods.• Currently working on novel exploration methods for target-driven 6-DOF grasping in highly cluttered scenes.Select Publications• “On-Policy Dataset Synthesis for Learning Robot Grasping Policies Using Fully Convolutional Deep Networks", IEEE Robotics & Automation Letters (2019) and IEEE International Conference on Robotics and Automation (2019)• “Learning ambidextrous robot grasping policies”, Science Robotics 4.26 (2019)• "Deep learning can accelerate grasp-optimized motion planning", Science Robotics 5.48 (2020)

• Worked on distributed reinforcement learning algorithms.• Worked on model parallelism in TensorFlow 2.0 using the Ray (https://github.com/ray-project/ray) distributed computing framework.• Worked on resource-constrained population based training of machine learning models.

• Sourced industry project collaborations and managed project teams.

• Led and advised teams of fellow undergraduate students in developing machine learning models for recommendation systems, anomaly detection, and addressing geo-diversity.