Researchers at NVIDIA and the University of Washington trained one simulated AI controller across all 20 assembly tasks on NIST’s Assembly Task Board #1. Their paper reports 98.2% overall training success. That is an aggregate result across the tasks, not a claim that every attempt succeeded.
Simulation frames from a ten-component assembly sequence.
The benchmark’s 20 tasks are distinct assembly pairs: each asks the robot to fit one part to its matching geometry. They include peg insertion, gear meshing, electrical-connector insertion and nut threading. The researchers trained one policy, or controller, to handle the different task types without demonstrations, hand-written motion routines or per-task tuning.
The policy received the robot’s state and point-cloud descriptions: sets of 3-D points representing the simulated arm, the loose part and the geometry it had to fit. It did not receive a task ID. The task reward arrived only when the assembly was within 1 millimeter of its target; a separate per-step cost discouraged large actions, high joint speeds and excess torque, without telling the robot whether it was getting closer.
For comparison, the authors trained one shared policy across the 20 tasks and 20 specialist policies, one for each task. The authors say each run had the same interaction allowance, but do not report how many interactions it contained. The shared policy reached 98.2% training success using 410 GPU-hours; the specialists averaged 81.4% and used 16,100 GPU-hours in total across 20 runs. So the matched budget was per run, not across the whole specialist group. The shared-policy setup used 64,000 parallel simulated environments and 25,000 iterations.
A separate experiment tested episodes with up to ten active assembly pairs. The robot could leave one part staged, assemble another, then return; reward came only when every active pair was complete. At ten pairs, reported success was 84%. The paper describes the staged behavior as planning-like coordination. This experiment involved at most ten simultaneous pairs; it does not mean the robot assembled all 20 benchmark tasks in one episode.
All results are in simulation. The authors used a simulated Franka Panda arm with hardware-matched speed and torque limits, but say this does not establish robustness to perception errors, delays or differences between simulated and real contact. Testing on the physical board remains future work. Extending the separate experiment from at most ten simultaneously active pairs to the board's 20 movable parts remains future work; the paper reports no 20-part episode.
Image credit: Figure 1 from AlphaNIST: Emergent Assembly Skills from Binary Task Rewards, by Octi Zhang et al.; CC BY 4.0.
Sources