How can robots acquire skills through interactions with the physical world? An interview with Jiaheng Hu - Robohub

The article features an interview with Jiaheng Hu discussing their research on enabling robots, particularly household mobile manipulators, to autonomously acquire skills through real-world reinforcement learning (RL). Traditional RL approaches in robotics often rely on training policies entirely in high-fidelity simulations before deploying them in the real world (zero-shot sim2real). However, this approach faces significant challenges: creating accurate, task-specific simulators is time-consuming and some tasks involving complex interactions (e.g., pouring water, folding clothes) are difficult to simulate realistically. Real-world RL, where robots learn directly through physical interaction, offers a promising alternative but is hindered by sample inefficiency and safety risks during exploration. To address these challenges, Hu and colleagues developed SLAC (Simulation-Pretrained Latent Action Space for Whole-Body Real-World RL), a two-step method that leverages low-fidelity simulation to facilitate safer and more efficient real-world learning. First, SLAC uses unsupervised RL in simulation to learn a