Papers
28
Total Citations
815
H-Index
13
About
Pulkit Agrawal is a prominent robotics and machine learning researcher whose work spans robotic manipulation, locomotion, imitation learning, and object representation. At the intersection of deep learning and physical interaction, Agrawal has made foundational contributions to how robots learn to understand and act upon the world around them. Among his most influential contributions is the development of Neural Descriptor Fields (138 citations), which introduced SE(3)-equivariant object representations enabling robots to generalize manipulation skills across object categories. His early work on experiential learning of intuitive physics — where a robot autonomously accumulated over 400 hours of poking experience — demonstrated the power of self-supervised, data-driven approaches to physical reasoning (132 citations). His research on rapid locomotion via reinforcement learning achieved record-breaking speeds of 3.9 m/s for the MIT Mini Cheetah (116 citations), while DribbleBot extended agile legged robotics to real-world dynamic manipulation tasks. Agrawal has also advanced imitation learning through zero-shot visual imitation and data-efficient frameworks like JUICER, addressing the persistent challenge of learning from limited demonstrations. Collectively, his body of work — spanning hundreds of citations — reflects a consistent drive to make robots more capable, adaptable, and practical in unstructured real-world environments.
Research Focus
Key Achievements
Top Papers
- 1
- 2Learning to Poke by Poking: Experiential Learning of Intuitive Physics132 citations · 2016
- 3Rapid Locomotion via Reinforcement Learning116 citations · 2022
- 4Zero-Shot Visual Imitation73 citations · 2018
- 5
- 6
- 7DribbleBot: Dynamic Legged Manipulation in the Wild43 citations · 2023
- 83D Neural Scene Representations for Visuomotor Control26 citations · 2021
- 9JUICER: Data-Efficient Imitation Learning for Robotic Assembly21 citations · 2024
- 10