AlphaNIST solves all 20 NIST assembly tasks in sim
Octi Zhang says one policy, trained from scratch on a pass or fail reward, hits 98.2% across the full NIST benchmark.
Octi Zhang released AlphaNIST, a single policy that he says solves all 20 NIST assembly tasks at a 98.2% success rate in simulation. The NIST tasks are a standard set of assembly problems, pegs, gears, connectors and the rest, used to measure robot manipulation. One policy covers all of them.
The part people are reacting to is how it was trained. No demonstrations, no shaped reward engineered per task. Zhang says it starts tabula rasa with a binary reward for success or failure plus a simple control cost, the same setup that drove AlphaGo and AlphaZero. He says the inspiration was direct, the question of what a robot would discover if given that same freedom.
Ten tasks, one reward at the end
The harder claim is sequencing. Zhang says AlphaNIST can complete 10 different tasks in a row while receiving the reward only once, after all 10 are done. That is a brutal credit assignment problem, the policy has to figure out which of hundreds of actions across ten assemblies mattered, from a single bit of feedback at the end.
He also says the policy found dexterous strategies the team could not have come up with themselves, the same thing AlphaGo's creators said about their system.
Shown versus claimed
The 98.2% is simulation only. Nothing in the posts reports a real-robot number, so whether any of this survives contact with real hardware is unknown right now. That matters because assembly is exactly where simulation tends to lie, contact physics and tight tolerances are the parts simulators model worst.
That did not stop people in the field from treating it as a sim-to-real story anyway. Simon Kalouche called it the "First promising result for sim2real in manipulation". Humanoid Scott's read was shorter, that dexterity lies more in the head than it does in the hand, a nod to the fact that nothing here is about a better gripper.
First promising result for sim2real in manipulation
Dexterity lies more in the head than it does in the hand
The paper and project website are linked from Zhang's post. The thing worth watching is the first real-robot success rate, and nobody has posted one yet.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.