Aviral Kumar
Papers
19
Total Citations
991
H-Index
8
About
Aviral Kumar is a prominent researcher at the forefront of offline reinforcement learning (RL) and robot learning, whose work has fundamentally shaped how AI systems leverage pre-collected data to acquire complex behaviors. His landmark 2020 tutorial, "Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems," has garnered nearly 800 citations and remains an essential reference for researchers entering the field, establishing the conceptual foundations for algorithms that learn exclusively from static datasets without costly online interaction. Kumar's research spans theoretical and applied dimensions of offline RL, including critical questions about when offline RL outperforms behavioral cloning, how to rigorously evaluate off-policy methods, and how pre-training on prior robotic experience can dramatically accelerate learning of new tasks. His contributions to scalable approaches — including Q-Transformer, which applies autoregressive transformers to offline RL — push the boundaries of multi-task robot learning. He has also advanced offline model-based optimization through benchmark tools like Design-Bench, broadening impact into domains such as protein design and computational engineering. Across robotics applications, Kumar consistently demonstrates how intelligently reusing past data reduces real-world collection burdens, making autonomous robot skill acquisition increasingly practical and accessible.
Research Focus
Key Achievements
Top Papers
- 1
- 2
- 3
- 4Benchmarks for Deep Off-Policy Evaluation25 citations · 2021
- 5
- 6
- 7Conservative Objective Models for Effective Offline Model-Based Optimization10 citations · 2021
- 8
- 9
- 10