Applying RL to Real-World Robotics with Abhishek Gupta - #466

Reinforcement learning (RL) in robotics faces significant hurdles due to mismatched assumptions between controlled simulations and unpredictable real-world environments. Abhishek Gupta, a PhD student at UC Berkeley, addresses these challenges by developing methods for autonomous, reset-free learning that function without manual instrumentation. Key strategies include unsupervised skill discovery, where robots optimize for diverse behaviors to prepare for future tasks, and "gradient surgery" to mitigate interference during multitask learning. By shifting focus toward "ecological RL," the research explores how dynamic, human-centric environments can naturally provide the helpful interventions necessary for robots to practice and refine skills. Moving beyond laboratory constraints, these approaches aim to scale robotic learning by enabling systems to collect data autonomously in real-world settings, ultimately bridging the gap between high-level algorithmic potential and physical execution.