Marvin Zhang

About

Marvin Zhang is a leading researcher at the intersection of deep reinforcement learning (RL) and robotics, whose work has fundamentally advanced how robots learn complex, real-world behaviors. His core contributions span model-based RL, policy learning under partial observability, and data-efficient robot control. Zhang is best known for his pioneering work on "SOLAR" (131 citations), which introduced deep structured representations for model-based RL, enabling data-efficient learning directly from high-dimensional image observations. He also made significant strides in combining model-based and model-free updates for trajectory-centric RL (88 citations), a hybrid approach that proved critical for real-world robotic applications. For partially observed control tasks, Zhang developed methods for learning policies with continuous memory states (90 citations), allowing robots to remember salient information from past observations—a key capability for complex manipulation. His work on tensegrity robot locomotion (92 citations) demonstrated deep RL's potential for controlling novel, compliant structures, with applications to planetary exploration rovers. More recently, his "AVID" framework (31 citations) tackled multi-stage task learning through pixel-level translation of human demonstrations, reducing the human burden in robotic RL. With over 450 total citations, Zhang's research continues to shape the future of intelligent, autonomous robotic systems.

Research Focus

Key Achievements

Top Papers

1. 1
   SOLAR: Deep Structured Representations for Model-Based Reinforcement Learning

   131 citations · 2018
2. 2
   Deep reinforcement learning for tensegrity robot locomotion

   92 citations · 2017
3. 3
   Learning deep neural network policies with continuous memory states

   90 citations · 2016
4. 4
   Combining Model-Based and Model-Free Updates for Trajectory-Centric Reinforcement Learning

   88 citations · 2017
5. 5
   AVID: Learning Multi-Stage Tasks via Pixel-Level Translation of Human Videos

   31 citations · 2020
6. 6
   Policy Learning with Continuous Memory States for Partially Observed Robotic Control.

   11 citations · 2015
7. 7
   Deep Reinforcement Learning for Tensegrity Robot Locomotion

   11 citations · 2016
8. 8
   Learning Deep Neural Network Policies with Continuous Memory States

   8 citations · 2015
9. 9
   AVID: Learning Multi-Stage Tasks via Pixel-Level Translation of Human\n Videos

   2 citations · 2019

Key Collaborators

Contact & Links