Training robots via reinforcement learning is expensive because each physical trial is costly and only reveals one outcome path. WorldSample addresses this by combining real robot rollouts with a learned "world model" that generates additional high-fidelity synthetic experience, reducing the need for physical trials. A technique called Policy-Paced Learning carefully schedules how synthetic data is used to avoid compounding hallucination errors or overestimating value. On contact-rich manipulation tasks, WorldSample boosts success rates by 28% while cutting training steps by 59%. This approach could substantially lower the cost and time required to deploy capable RL-trained robots in real-world settings.
Authors: Yuquan Xue, Le Xu, Zeyi Liu, Zhenyu Wu, Zhengyi Gu, Xinyang Song, Bofang Jia, Ziwei Wang
Paper: https://arxiv.org/abs/2607.02431v1
Weekly, digestible podcast explainers of significant research papers