Arunkumar Byravan

About

Arunkumar Byravan is a robotics and machine learning researcher whose work spans deep reinforcement learning, structured dynamics models, and visuomotor control, with a particular focus on enabling robots to learn complex physical skills from data. His most celebrated contribution — teaching a bipedal humanoid robot to play one-versus-one soccer using deep RL — has already garnered 147 citations since 2024, demonstrating the field's excitement around agile, learned locomotion in low-cost hardware. His SE3-Nets and SE3-Pose-Nets frameworks brought elegant geometric structure to deep dynamics modeling, allowing robots to segment and predict rigid-body motion directly from raw point cloud and depth data — work that continues to influence robot perception and planning. His NeRF2Real system cleverly bridges the sim-to-real gap by reconstructing real-world scenes from phone video for visually grounded bipedal control. Earlier contributions in functional gradient motion planning and inverse optimal control from demonstration reflect a career-long interest in principled, efficient robot decision-making. More recently, Byravan has explored foundation models as unified robotic agents, signaling ambitions toward generalizable, language-grounded robot intelligence. Across his portfolio, his research consistently advances the frontier of physically capable, perceptually aware autonomous systems.

Research Focus

Key Achievements

Top Papers

1. 1
   Learning agile soccer skills for a bipedal robot with deep reinforcement learning

   147 citations · 2024
2. 2
   SE3-Pose-Nets: Structured Deep Dynamics Models for Visuomotor Control

   44 citations · 2018
3. 3
   NeRF2Real: Sim2real Transfer of Vision-guided Bipedal Motion Skills using Neural Radiance Fields

   43 citations · 2023
4. 4
   Space-time functional gradient optimization for motion planning

   37 citations · 2014
5. 5
   Learning predictive models of a depth camera & manipulator from raw execution traces

   35 citations · 2014
6. 6
   SE3-Pose-Nets: Structured Deep Dynamics Models for Visuomotor Planning and Control

   31 citations · 2017
7. 7
   SE3-nets: Learning rigid body motion using deep neural networks

   27 citations · 2017
8. 8
   Graph-based inverse optimal control for robot manipulation

   18 citations · 2015
9. 9
   Towards A Unified Agent with Foundation Models

   17 citations · 2023
10. 10
    Beyond Pick-and-Place: Tackling Robotic Stacking of Diverse Shapes

    16 citations · 2021

Key Collaborators

Contact & Links