Reinforcement Learning featuring Vikash Kumar

In this talk

1. 2:08

   Why robotics is turning to real-world data

   Vikash says progress comes from both richer simulation and better data-driven decision making. Simulations can miss details such as friction, material behavior, and folds, which matter for delicate tasks like spinning a pen. Researchers can use real-world experience to improve models, or combine simulation with data to close the gap between them.

2. 7:22

   ROBEL lowers barriers to research

   Vikash describes ROBEL as an effort to make robotics more affordable and usable beyond carefully controlled settings. Its D’Claw, D’Lantern, and D’Kitty platforms let more researchers experiment and gather real-world data. He says learning systems need robots that are repeatable, observable, and scalable, even if they lack the precision of costly industrial arms.

3. 13:10

   Scale, safety, and useful robot applications

   Vikash says learning needs more data, so progress depends both on collecting more safely and on reducing how much data algorithms require. He expects legged robots to find useful roles in risky settings, with meaningful impact potentially emerging in two to five years; manipulation faces higher safety demands. Wheels suit smooth routes and large buildings, while legs may handle rough sidewalks better.

4. 30:02

   Continual learning in a changing world

   Vikash explains that skills learned in one setting may stop working as conditions change. A parked car, for example, may face different surroundings when its driver returns, requiring adaptation. He frames continual learning as updating useful skills, retiring outdated ones, and combining existing abilities as environments and other agents change.

5. 38:46

   Reliable hardware and getting started

   After a costly, specialized hand project, Vikash sought systems that other researchers could reproduce and run continuously. He says modular Dynamixel-based robots operated around the clock and accumulated more than five years of data, making real-world experimentation practical. His advice to newcomers is to start tinkering, expect to break things, and learn through rebuilding.