AlphaNIST demonstrates that assembly skills can emerge when only a binary task reward is provided on the NIST Assembly Task Board #1. The benchmark demands precise, contact‑rich control of small, tight‑fitting parts over short horizons and recovery and coordination across long action sequences. Prior work relied on demonstrations, task decomposition, or dense progress rewards to guide assembly, which implicitly prescribe strategy. AlphaNIST asks whether a sparse success/failure reward, together with task diversity, can let a robot discover how to assemble on its own. In simulation we train a single policy from scratch across all 20 assembly tasks, achieving a 98.2% overall success rate without demonstrations, progress shaping, scripted primitives, or per‑task tuning. The framework also trains policies to complete ten assembly pairs in one episode, rewarding success only when all ten pairs are simultaneously finished, with no prescribed order. The learned policies exhibit planning‑like behaviors: selecting grasps that facilitate later assembly, using surrounding geometry to reorient parts, preparing part A, assembling part B, then returning to finish part A, all without an explicit planner. See videos at https://alpha-nist.github.io/.
Review