Reinforcement-learning

Researchers from MIT's Improbable AI Lab have developed SoftMimic, a reinforcement learning framework that trains humanoids to be compliant and safely absorb collisions, rather than rigidly tracking motions.

Amazon's latest robotics project, ResMimic, uses a two-stage residual learning framework to efficiently teach humanoids complex loco-manipulation skills. By refining a general motion policy with task-specific corrections, the system enables a Unitree G1 to handle heavy and irregular objects with precision.

A day after unveiling its Redwood AI brain, robotics firm 1X has detailed the reinforcement learning system that powers the NEO humanoid’s movement. The unified controller enables a range of human-like motions, from walking and running to climbing stairs using only stereo vision.