Satvik Sharma

I am a Master's student at Berkeley majoring in Electrical Engineering and Computer Science. I currently work in Berkeley Artificial Intelligence Research (BAIR) and the AUTOLAB where I am advised by Prof. Ken Goldberg. I also did my bachelors degree at Berkeley in CS. I have also collaborated with Pieter Abbeel, Angjoo Kanazawa, Jitendra Malik, Anca Dragan, and researchers from Google DeepMind.

Email  /  Google Scholar  /  Github  /  Linkedin  

My research vision is enabling robots to perform useful long-horizon tasks without significant task-specific online robot data. I am excited about leveraging methods from vision, language, and theory to make robot learning efficient. My research has focused on developing sample-efficient methods for robotic perception and robustness to uncertainty, both important for learning long-horizon tasks such as deformable manipulation, object search, and grasping. Currently, I'm interested in developing algorithms for imitation learning, multi-modal robot learning, and robot perception specfically for useful, long-horizon tasks.

Modular framework explicitly separating VLMs and LLMs for occluded object search, which outperforms CLIP-based methods.

LERF-TOGO uses CLIP and DINO features for language-specified tasks performing zero-shot semantic grasping.

A scalable and robust RL algorithm which optimizes for a combination of expected performance and tail risk under a distribution over learned reward functions.

We introduce new formalism, algorithms, and open-source benchmarks for "Interactive Fleet Learning": interactive learning with multiple robots and multiple humans. .

We study strategies to automatically determine when policies trained in simulation can be reliably transferred to a physical robot, specifically for a fabric smoothing task.

Planning multi-object grasps under frictional constraints using a grasp success predictor network trained in real. We increase picks per hour over baselines and are robust to grasp and state uncertainity.

We study how we can use pretrained ensembles - encouraged via a KL-divergence in their loss function to be as diverse as possible - to then generalize to new tasks using RL.

Comparing the AlphaGarden system (simulator and Farmbot actuation) against professional horticulturalists in terms of plant diversity and canopy coverage.

Fully autonomous system for pruning polyculture plants with specialized vision algorithms for identifying plants and tracking their centers.