FastSAC: Humanoid Locomotion in 15 Minutes скачать в хорошем качестве

FastSAC: Humanoid Locomotion in 15 Minutes

In this AI Research Roundup episode, Alex discusses the paper: 'Learning Sim-to-Real Humanoid Locomotion in 15 Minutes(2512.01996v1)' This work shows how massively parallel simulation and off-policy reinforcement learning can train robust humanoid locomotion controllers in just 15 minutes on a single RTX 4090 GPU. The authors introduce a practical recipe based on FastSAC and FastTD3 that stabilizes off-policy learning at scale with thousands of parallel environments. With minimalist reward functions and strong domain randomization, their controllers handle randomized dynamics, rough terrain, and push perturbations. They demonstrate results on Unitree G1 and Booster T1 robots, as well as fast whole-body human-motion tracking, and release open-source implementations and videos. Paper URL: https://arxiv.org/pdf/2512.01996 #AI #MachineLearning #DeepLearning #ReinforcementLearning #Robotics #HumanoidRobots #Sim2Real