Visak Kumar

About

Visak Kumar is a robotics researcher whose work sits at the intersection of reinforcement learning and real-world robot control, with a particular focus on bridging the "sim-to-real" gap. His most influential contribution, "Sim-to-Real Transfer for Biped Locomotion" (25 citations), introduced a novel two-stage system identification approach that enables dynamic policies trained in simulation to transfer effectively to physical hardware—a critical challenge in legged robotics. Kumar has also advanced dexterous manipulation through "Contextual Reinforcement Learning of Visuo-tactile Multi-fingered Grasping Policies" (12 citations), integrating tactile sensing with vision to train robust grasping policies. His work on "Improving Model-Based Balance Controllers Using Reinforcement Learning and Adaptive Sampling" (8 citations) combines classical control theory with data-driven methods to enhance humanoid balance recovery. More recently, in "Joint Space Control via Deep Reinforcement Learning" (4 citations), he proposed a model-free neural network controller that replaces traditional inverse kinematics, offering a simpler, versatile alternative for manipulator control. Across these contributions, Kumar demonstrates a consistent commitment to making simulated training practical for complex, real-world robotic systems.

Research Focus

Key Achievements

Top Papers

1. 1
   Sim-to-Real Transfer for Biped Locomotion

   25 citations · 2019
2. 2
   Contextual Reinforcement Learning of Visuo-tactile Multi-fingered\n Grasping Policies

   12 citations · 2019
3. 3
   Improving Model-Based Balance Controllers Using Reinforcement Learning and Adaptive Sampling

   8 citations · 2018
4. 4
   Joint Space Control via Deep Reinforcement Learning

   4 citations · 2021

Key Collaborators

Contact & Links