Todd Hester

About

Todd Hester is a prominent researcher specializing in reinforcement learning (RL) for robotics, with particular expertise in sample-efficient learning, safe exploration, and learning from demonstrations. His work addresses some of the most pressing challenges in applying RL to real-world physical systems, where data collection is costly and operational constraints are critical. Hester's most influential contribution, "Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards" (510 citations), introduced a landmark approach that combines human demonstrations with deep RL to dramatically accelerate learning under sparse reward conditions — a persistent bottleneck in robotic applications. Complementing this, his work on safe exploration (275 citations) tackled the vital problem of deploying RL agents in high-stakes environments, such as datacenter cooling systems, without violating critical operational constraints. His earlier TEXPLORE and RTMBA frameworks demonstrated innovative model-based RL architectures capable of real-time, sample-efficient learning on humanoid robots, establishing him as a pioneer in practical robot learning. His research on intrinsic motivation and robustness to model misspecification further broadens his impact across the field. Beyond research, Hester contributed to championship-level robot soccer as part of the UT Austin Villa team, showcasing the real-world applicability of his methods.

Research Focus

Key Achievements

Top Papers

1. 1
   Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards

   510 citations · 2017
2. 2
   Safe Exploration in Continuous Action Spaces

   275 citations · 2018
3. 3
   TEXPLORE: real-time sample-efficient reinforcement learning for robots

   104 citations · 2012
4. 4
   Generalized model learning for Reinforcement Learning on a humanoid robot

   96 citations · 2010
5. 5
   RTMBA: A Real-Time Model-Based Reinforcement Learning Architecture for robot control

   86 citations · 2012
6. 6
   Intrinsically motivated model learning for developing curious robots

   83 citations · 2015
7. 7
   Robust Reinforcement Learning for Continuous Control with Model Misspecification

   38 citations · 2019
8. 8
   Negative information and line observations for Monte Carlo localization

   25 citations · 2008
9. 9
   TEXPLORE: Temporal Difference Reinforcement Learning for Robots and Time-Constrained Domains

   13 citations · 2013
10. 10
    UT Austin Villa 2012: Standard Platform League World Champions

    11 citations · 2013

Key Collaborators

Contact & Links