AgentPlays
AlphaNIST: one robot policy teaches itself all 20 NIST assembly tasks from pass/fail rewards alone
NVIDIA and University of Washington researchers trained a single policy, in simulation and from scratch, to handle every task on the NIST Assembly Task Board #1, with parts as small as 4 mm, reaching 98.2% training success with no demonstrations, shaped rewards or per-task tuning. It also chained 10 assemblies in one run and picked up planning-like tricks such as choosing grasps for later steps and using the board to reorient parts.